diff --git a/.gitattributes b/.gitattributes index e7c9b02859d2f5b70d33618e7ed3ad74283d7286..7f45d6d70a8c5ad84effe53b1cea7c169b4c99aa 100644 --- a/.gitattributes +++ b/.gitattributes @@ -101,3 +101,74 @@ hyperdetailed-colored-pencil/example_images/03_image.jpeg filter=lfs diff=lfs me hyperdetailed-colored-pencil/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text hyperdetailed-colored-pencil/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text hyperdetailed-colored-pencil/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +slavekini-aka-slave-leia-outfit-clothing-krea2-anima-il/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +slavekini-aka-slave-leia-outfit-clothing-krea2-anima-il/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +slavekini-aka-slave-leia-outfit-clothing-krea2-anima-il/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +slavekini-aka-slave-leia-outfit-clothing-krea2-anima-il/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +slavekini-aka-slave-leia-outfit-clothing-krea2-anima-il/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +slavekini-aka-slave-leia-outfit-clothing-krea2-anima-il/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +bartstyle-blue-archive-art-style-lora-anime-flat-color/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +bartstyle-blue-archive-art-style-lora-anime-flat-color/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +bartstyle-blue-archive-art-style-lora-anime-flat-color/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +bartstyle-blue-archive-art-style-lora-anime-flat-color/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +bartstyle-blue-archive-art-style-lora-anime-flat-color/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +bartstyle-blue-archive-art-style-lora-anime-flat-color/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +2000s-analog-core/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +2000s-analog-core/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +2000s-analog-core/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +2000s-analog-core/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +2000s-analog-core/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +2000s-analog-core/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +granblue-fantasy-style/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +granblue-fantasy-style/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +granblue-fantasy-style/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +granblue-fantasy-style/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +granblue-fantasy-style/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +granblue-fantasy-style/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +famegrid-krea-2-z-image-qwen-realism-lora/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +famegrid-krea-2-z-image-qwen-realism-lora/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +famegrid-krea-2-z-image-qwen-realism-lora/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +famegrid-krea-2-z-image-qwen-realism-lora/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +famegrid-krea-2-z-image-qwen-realism-lora/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +famegrid-krea-2-z-image-qwen-realism-lora/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +the-space-marines-warhammer-40k-flux-pony-illustrious-anima/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +the-space-marines-warhammer-40k-flux-pony-illustrious-anima/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +the-space-marines-warhammer-40k-flux-pony-illustrious-anima/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +the-space-marines-warhammer-40k-flux-pony-illustrious-anima/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +the-space-marines-warhammer-40k-flux-pony-illustrious-anima/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +the-space-marines-warhammer-40k-flux-pony-illustrious-anima/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +krea2-textfusion-refusal-reduction-lora/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +krea2-textfusion-refusal-reduction-lora/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +krea2-textfusion-refusal-reduction-lora/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +krea2-textfusion-refusal-reduction-lora/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +krea2-textfusion-refusal-reduction-lora/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +krea2-textfusion-refusal-reduction-lora/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +background-detail-enhancer/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +background-detail-enhancer/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +background-detail-enhancer/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +background-detail-enhancer/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +background-detail-enhancer/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +background-detail-enhancer/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +aesthetic-amateur-photo/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +aesthetic-amateur-photo/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +aesthetic-amateur-photo/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +aesthetic-amateur-photo/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +aesthetic-amateur-photo/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +aesthetic-amateur-photo/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +ultrareal-krea2-klein9b/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +ultrareal-krea2-klein9b/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +ultrareal-krea2-klein9b/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +ultrareal-krea2-klein9b/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +ultrareal-krea2-klein9b/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +ultrareal-krea2-klein9b/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +new-fantasy-core-ill-flux-pony-sdxl-zit-detailer/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +new-fantasy-core-ill-flux-pony-sdxl-zit-detailer/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +new-fantasy-core-ill-flux-pony-sdxl-zit-detailer/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +new-fantasy-core-ill-flux-pony-sdxl-zit-detailer/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +new-fantasy-core-ill-flux-pony-sdxl-zit-detailer/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +new-fantasy-core-ill-flux-pony-sdxl-zit-detailer/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +satoshi-urushihara-style/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +satoshi-urushihara-style/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +satoshi-urushihara-style/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +satoshi-urushihara-style/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +satoshi-urushihara-style/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text diff --git a/2000s-analog-core/example_images/01_image.jpeg b/2000s-analog-core/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..556f39ab07c0d04b7a5666856fdc4a706e539eaf --- /dev/null +++ b/2000s-analog-core/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fcdd83c8f68572bcdf5df3568e09039428ac923817957fc0a5f7bd5ad2e5aaa0 +size 1181608 diff --git a/2000s-analog-core/example_images/02_image.jpeg b/2000s-analog-core/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..4f8af392261d0e19719b55f167c5b56a58d873c1 --- /dev/null +++ b/2000s-analog-core/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c7c085429354e73a53d35c0d84d8a6bd0a5dfacfa6ac2c0b10768f4973f78448 +size 1526585 diff --git a/2000s-analog-core/example_images/03_image.jpeg b/2000s-analog-core/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..6aec7e6cb50e39144854967bcbc99eab28a8ccce --- /dev/null +++ b/2000s-analog-core/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:00322b847c0350f793756456a2d7404e1f0d49866c6fbd99030084a2af36eba5 +size 1410849 diff --git a/2000s-analog-core/example_images/04_image.jpeg b/2000s-analog-core/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..58033230ca1a96fc93b143162b9c24a139c68fe4 --- /dev/null +++ b/2000s-analog-core/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3537c7af07c62af4ca034c8b780441a7b06a9dc4379fb23b76b16b35c121e6b5 +size 2099590 diff --git a/2000s-analog-core/example_images/05_image.jpeg b/2000s-analog-core/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..125a89a16c6442cf628cbcb0474290215fb60d0f --- /dev/null +++ b/2000s-analog-core/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6fb1dcddf345184e9c70061365a4870ae12f4185b863a4ab7fa1c920361c06a6 +size 1271734 diff --git a/2000s-analog-core/example_images/06_image.jpeg b/2000s-analog-core/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..e11f4a8413520c129c2f98a769202b660713ee90 --- /dev/null +++ b/2000s-analog-core/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ca70eeb8215b626cc645f2d5ad8a71d61704e5509b1f7406efdb6fe8f4db3438 +size 1415484 diff --git a/2000s-analog-core/info.txt b/2000s-analog-core/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..bdd0c1f6cff023b167a5f3ba065541a203075a8b --- /dev/null +++ b/2000s-analog-core/info.txt @@ -0,0 +1,311 @@ +═══════════════════════════════════════════════════════════════════════════════ + 2000S ANALOG CORE +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Visual style + Author Danrisi + Source https://civitai.com/models/1134895 + Provenance CivitAI · model 1134895 · version 3103914 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File vhs_krea2.safetensors + Version v1.0 Krea2 + Size 109.0 MB + Format SafeTensor / unknown precision + SHA256 cef79e53f88941e29ea7211543f624981207e32f15abef80c56fea3d0676e533 + Published 2026-07-06 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + No trigger word. The LoRA applies as soon as it is loaded — + strength is the only way to dial the effect up or down. + +─────────────────────────────────────────────────────────────────────────────── + RECOMMENDED SETTINGS +─────────────────────────────────────────────────────────────────────────────── + Taken verbatim from the author's description: + + LoRA strength 0.9 + Sampler steps 40 + Resolution 704×704 + +─────────────────────────────────────────────────────────────────────────────── + ADOPTION & POPULARITY +─────────────────────────────────────────────────────────────────────────────── + Downloads 30,523 + Rating 100% positive (2334/2336) + Comments 0 + Published 60 days ago + Download rate 89.9/day since release + + Adoption rank #18 of 150 in this collection (High) + Momentum rank #30 of 150 by download rate (Trending) + + Adoption tier is the percentile of total downloads within this collection; + momentum is the percentile of downloads-per-day since release. Momentum is + ranked rather than measured against a fixed rate, because this base model + is itself new and every LoRA looks fast on an absolute scale. Both numbers + are CivitAI's, read on the scrape date at the foot of this file. + +─────────────────────────────────────────────────────────────────────────────── + CHAINING / STACKING +─────────────────────────────────────────────────────────────────────────────── + Stacks with others not stated + + The author gave no chaining guidance. The rules below are the + general Krea 2 ones, not anything specific to this LoRA. + + General Krea 2 rules that apply here (see REFERENCE.txt): + + • The most important rule is not about chaining but about which + checkpoint you chain onto: the official README says in bold "TRAIN on + Raw and RUN on Turbo", and "We highly recommend using RAW for training + LoRAs and applying them on Turbo for inference." LoRAs trained on Raw + are stated to work well on Turbo. + • Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 + LoRAs are DiT-only, so there is no CLIP side to chain and the text- + encoder branch stays untouched however many LoRAs you add. + • Krea 2 LoRAs stack in practice and the community treats multi-LoRA as + routine - style + character, style + slider, realism + character. + Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one + writes "I recently stacked over 20 of my own LoRAs on a single + generation and the output still came out great." + • Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge + on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is + loaded. The CivitAI 12GB training guide also notes rank affects + composability: "rank 8 produced a lighter style influence that was + easier to stack with other LoRAs, rank 16 provided a useful balance + between concept strength and composability, and rank 32 was better + suited to a precise object or concept intended to take priority over + other stacked LoRAs." + • Accelerator LoRAs are position-sensitive in intent, not in node order: + a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a + sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to- + Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is + explicitly recommended by the 4-step LoRA's author. + • Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI + (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card + says "Keep mu = 1.15. The training targets are anchored to that grid; a + different shift moves the [sampler off them]"; the same reasoning + applies to any Turbo-trained adapter in the chain. + • Reference/edit LoRAs are designed to be stacked under ordinary subject + and style LoRAs - the Identity Edit card states "Composes with your + LoRAs: character/body/style LoRAs stack on top and steer the prior" and + recommends a stacked subject LoRA when unusual subjects drift toward + the base prior. + • Krea 2 LoRAs cannot be chained with LoRAs for any other base model. + Many CivitAI pages carry multi-model titles like "(Krea-2 + ZIT)" or + "[Flux | ZIB | Krea2]" - those are separate files for separate + architectures shipped under one model page, not a combined adapter. + • To combine two full Krea 2 checkpoints rather than adapters, ComfyUI + ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per- + section weights for first., tmlp., txtmlp., tproj., + txtfusion.projector, txtfusion.layerwise_blocks.0-1, + txtfusion.refiner_blocks.0-1, blocks.0-27 and last. + • The author never states which weight family this targets. Silent + degradation, never a hard error, between Raw and Turbo. The two share + one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" + - so every LoRA tensor shape matches on both and ComfyUI/diffuser + +─────────────────────────────────────────────────────────────────────────────── + PROS +─────────────────────────────────────────────────────────────────────────────── + + Widely adopted — 30,523 downloads. [source: CivitAI stats] + + Positively received — 2,334 thumbs-up, 2 down. [source: CivitAI stats] + + Compact at 109 MB — cheap to stack with other LoRAs. [source: file + size] + + Author published concrete recommended settings. [source: description] + +─────────────────────────────────────────────────────────────────────────────── + CONS & CAVEATS +─────────────────────────────────────────────────────────────────────────────── + - No trigger word — the LoRA is always on once loaded, so strength is the + only control. [source: model metadata] + - Author never states which generation mode it was trained for; test + against your own workflow before relying on it. [source: description] + +─────────────────────────────────────────────────────────────────────────────── + COMPATIBILITY WARNING +─────────────────────────────────────────────────────────────────────────────── + Silent degradation, never a hard error, between Raw and Turbo. The two + share one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" - so + every LoRA tensor shape matches on both and ComfyUI/diffusers will load an + adapter against either without complaint. The lvladikov 4-step LoRA card + states this for the reverse direction: "Every layer it targets also exists + in Krea 2 Raw, so it will load there without complaint - but that is a side + effect of the shared architecture, not a supported mode", and reports + mixed, subject-dependent results with malformed subjects (duplicated heads, + fused limbs) at low step counts, needing 14+ steps with Raw's CFG on before + output is coherent. The intended direction (train on Raw, run on Turbo) is + officially blessed. The more common practical failure is a settings + mismatch rather than a weight mismatch: running a Turbo-targeted LoRA at + Raw's 52 steps + CFG 3.5, or leaving CFG enabled on Turbo, degrades output + with no warning. Loading a Krea 2 LoRA against a different model family + entirely (Flux, Qwen-Image, Z-Image) is also not an error - ComfyUI logs + "lora key not loaded: {key}" per key and generates as if no LoRA were + applied. + +─────────────────────────────────────────────────────────────────────────────── + AUTHOR'S DESCRIPTION +─────────────────────────────────────────────────────────────────────────────── + Z-Image I used at strength 0.9 with res_2s_ode + beta. Workflow you can + parse from my images Qwen Edition: Here everything works perfectly, + timestamps, text - excellent. Images with women i generate with my new lora + https://civitai.com/models/1862761?modelVersionId=2338791 V3 Update: + Enhanced Timestamps + - Improved realism for on-screen display times and date stamps. + + - Face Fixes + - Reduced the repetition of identical faces across generated frames, + enhancing variety. + + - Expanded Dataset + - Now trained on VHS, VHS-C, and Hi8 footage to broaden the analog + aesthetic. + + Usage & Tips + - Low-Res VHS Look: + - Use prompts with phrases like “amateur quality” and “low resolution” to + push your render toward that classic, fuzzy VHS vibe. + + - Target ~0.5 megapixels (e.g. 704×704, 608 x 816) if you want reinforce + extreme lo-fi aesthetic. + + - Higher-Res Hi8 Style: + - To achieve the “Hi8” look (cleaner but still analog), render at ~1 + megapixel (e.g. 896 x 1152 or 1024×1024). + + - You can go up to ~2 megapixels if you want more detail without losing the + retro feel. + + - Recommended Trigger: + - The only trigger word you truly need for this model is “v8s” . + + - Avoid mixing too many triggers, as that can introduce confusion and stray + from the classic 2000s analog style. V2 Update: + + - Removed digital camera images from the dataset and added more footage + from camcorders. + + - Fixed the issue where the same woman appeared in all images. + + - Tried improving the sexy overlay text on footage, but it only generates + well sometimes. + + Can’t say I’m fully happy with this version, but it’s not bad. Already + planning a new update with more features soon! Also thinking about updating + the Mavica LoRA. + This keeps your style while improving clarity and readability. Let me know + if you want any tweaks P.S: but what i really like in this lora, that it's + for some reason generate almost always perfect hands Step back into the + early 2000s, where Myspace was king, profile pictures were taken with a + digital potato, and everything felt just a little bit off (in the best + way). This LoRA captures that lo-fi, analog aesthetic of forgotten VHS + tapes and grainy digital cameras (yeah, i mix them in one LoRa). Perfect + for those who miss the charm of pixelated edges and questionable lighting + choices. No, it won't resurrect your old Myspace profile - sadly - but it + will bring back all those nostalgic vibes. Warning: May cause a sudden urge + to listen to emo playlists P.S.S: only v8s is needed in trigger words, + everything else is optional to enhance effect Usage as usual: dpmpp2m + + beta + 40 steps + 2.5-3 guidance I'm using with UltraReal checkpoint , but + works good with default flux.dev too P.S: sometimes can be some timestamps + or other features from vhs tapes (will be fixed in next versions). I just + wanted to test some techniques on this LoRa and result is pretty solid for + 29 images in dataset. Also noticed one bug, that female faces often looks + similar, will fix in next release. But LoRas wit characters works good and + fix this problem + +─────────────────────────────────────────────────────────────────────────────── + EXAMPLE MEDIA +─────────────────────────────────────────────────────────────────────────────── + 6 preview(s) in ./example_images/ — community results for + this LoRA. Videos keep a matching _poster.jpg still frame. + + No prompts are recorded below: CivitAI's public API flags these posts as + having generation metadata but does not return it, so the prompts cannot be + scraped. Open the model page to read them. + + 01_image.jpeg (image, 1472x1024) + 02_image.jpeg (image, 1472x1024) + 03_image.jpeg (image, 1472x1024) + 04_image.jpeg (image, 1024x1472) + 05_image.jpeg (image, 1408x1088) + 06_image.jpeg (image, 1408x1088) + +─────────────────────────────────────────────────────────────────────────────── + HOW TO USE (COMFYUI) +─────────────────────────────────────────────────────────────────────────────── + 1. Update ComfyUI to at least v0.26.0 (released 2026-06-23), which added + local Krea2 support. For reference-image / style-reference / edit LoRAs you + need at least v0.29.0 (2026-07-29), which added regular and timestep-zero + reference images for Krea 2. ModelMergeKrea2 arrived in v0.27.0 + (2026-06-30). + + 2. Put the base stack in place (files from huggingface.co/Comfy-Org/Krea-2, + which is not gated): + - ComfyUI/models/diffusion_models/krea2_turbo_fp8_scaled.safetensors + (documented recommended default; the repo also holds krea2_turbo_bf16, + _int8_convrot, _mxfp8, _nvfp4 and krea2_raw_bf16 / _fp8_scaled / + _int8_convrot) + - ComfyUI/models/text_encoders/qwen3vl_4b_fp8_scaled.safetensors (bf16 + also available) + - ComfyUI/models/vae/qwen_image_vae.safetensors + + 3. Put the LoRA .safetensors in ComfyUI/models/loras/ - flat, no per-model + subfolder required. + + 4. Build or load the graph. The official image_krea2_turbo_t2i.json + template is: UNETLoader (krea2_turbo_fp8_scaled.safetensors) -> + LoraLoaderModelOnly -> KSampler; CLIPLoader + (qwen3vl_4b_fp8_scaled.safetensors, type "krea2") -> text encode -> + KSampler positive, with ConditioningZeroOut into negative; VAELoader + (qwen_image_vae.safetensors) -> VAEDecode -> SaveImage; EmptyLatentImage + 1024x1024 driven by ResolutionSelector. An int8 sibling template + (image_krea2_turbo_t2i_int8.json) also ships. + + 5. Load the LoRA with LoraLoaderModelOnly, not LoraLoader - both official + templates do. Krea 2 LoRAs patch the DiT only; note that the DiT's own + text_fusion / txtfusion blocks are a normal LoRA target (the official style + LoRAs patch them), but the Qwen3-VL text encoder itself is not. Set + strength 1.0 for Krea's official style LoRAs (the shipped template uses + 0.8) and add the LoRA's trigger word to the prompt. + + 6. Sampler settings: steps 8, cfg 1, sampler euler, scheduler simple, + denoise 1.0. Do not add a ModelSampling node for plain t2i - ComfyUI + already applies shift 1.15 from Krea2.sampling_settings. + + 7. For style reference, use the dedicated template: UNETLoader + krea2_turbo_int8_convrot.safetensors, LoraLoaderModelOnly + krea2_style_reference.safetensors at 1.0, ModelSamplingFlux (1.15, 0.5), + TextEncodeQwenImageEditPlus fed by LoadImage, + FluxKontextMultiReferenceLatentMethod set to index_timestep_zero, and + SamplerCustomAdvanced with KSamplerSelect euler + BasicScheduler simple/8 + + CFGGuider cfg 1. + + 8. For instruction editing (Krea 2 Identity Edit), stock nodes are not + enough: install the ComfyUI-Krea2Edit node pack, which ships its own + workflows, and run Turbo at 8-12 steps / cfg 1 (or Raw at 20 steps / cfg 3 + for removals) with LoRA strength 1.0. + + 9. Low-VRAM alternative: install city96's ComfyUI-GGUF and use a community + GGUF build (realrebelai/KREA-2_GGUFs, quantized from Turbo; + molbal/krea2-gguf, quantized from Raw) with the Unet Loader (GGUF) node in + place of UNETLoader. There is no official GGUF release. LoRA loading is + unchanged. + +─────────────────────────────────────────────────────────────────────────────── + Scraped by lorakit on 2026-09-05 00:11 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/2000s-analog-core/metadata.json b/2000s-analog-core/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..6c170a854261346b18ae058b42e2a5d422863d37 --- /dev/null +++ b/2000s-analog-core/metadata.json @@ -0,0 +1,141 @@ +{ + "name": "2000s Analog Core", + "slug": "2000s-analog-core", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 1134895, + "version_id": 3103914, + "version_name": "v1.0 Krea2", + "source": "civitai", + "url": "https://civitai.com/models/1134895", + "repo_id": null, + "file_path": null, + "author": "Danrisi", + "category": { + "key": "style", + "label": "Visual style" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [], + "recommended_settings": { + "LoRA strength": "0.9", + "Sampler steps": "40", + "Resolution": "704×704" + }, + "chaining": { + "author": { + "stacks": "not stated", + "notes": "", + "avoid": [], + "requires": [], + "order": "", + "strength": "", + "stated": false + }, + "general_rules": [ + "The most important rule is not about chaining but about which checkpoint you chain onto: the official README says in bold \"TRAIN on Raw and RUN on Turbo\", and \"We highly recommend using RAW for training LoRAs and applying them on Turbo for inference.\" LoRAs trained on Raw are stated to work well on Turbo.", + "Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 LoRAs are DiT-only, so there is no CLIP side to chain and the text-encoder branch stays untouched however many LoRAs you add.", + "Krea 2 LoRAs stack in practice and the community treats multi-LoRA as routine - style + character, style + slider, realism + character. Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one writes \"I recently stacked over 20 of my own LoRAs on a single generation and the output still came out great.\"", + "Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is loaded. The CivitAI 12GB training guide also notes rank affects composability: \"rank 8 produced a lighter style influence that was easier to stack with other LoRAs, rank 16 provided a useful balance between concept strength and composability, and rank 32 was better suited to a precise object or concept intended to take priority over other stacked LoRAs.\"", + "Accelerator LoRAs are position-sensitive in intent, not in node order: a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to-Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is explicitly recommended by the 4-step LoRA's author.", + "Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card says \"Keep mu = 1.15. The training targets are anchored to that grid; a different shift moves the [sampler off them]\"; the same reasoning applies to any Turbo-trained adapter in the chain.", + "Reference/edit LoRAs are designed to be stacked under ordinary subject and style LoRAs - the Identity Edit card states \"Composes with your LoRAs: character/body/style LoRAs stack on top and steer the prior\" and recommends a stacked subject LoRA when unusual subjects drift toward the base prior.", + "Krea 2 LoRAs cannot be chained with LoRAs for any other base model. Many CivitAI pages carry multi-model titles like \"(Krea-2 + ZIT)\" or \"[Flux | ZIB | Krea2]\" - those are separate files for separate architectures shipped under one model page, not a combined adapter.", + "To combine two full Krea 2 checkpoints rather than adapters, ComfyUI ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per-section weights for first., tmlp., txtmlp., tproj., txtfusion.projector, txtfusion.layerwise_blocks.0-1, txtfusion.refiner_blocks.0-1, blocks.0-27 and last.", + "The author never states which weight family this targets. Silent degradation, never a hard error, between Raw and Turbo. The two share one architecture - ai-toolkit's config is commented \"The reference 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)\" - so every LoRA tensor shape matches on both and ComfyUI/diffuser" + ] + }, + "pros": [ + "Widely adopted — 30,523 downloads. [source: CivitAI stats]", + "Positively received — 2,334 thumbs-up, 2 down. [source: CivitAI stats]", + "Compact at 109 MB — cheap to stack with other LoRAs. [source: file size]", + "Author published concrete recommended settings. [source: description]" + ], + "cons": [ + "No trigger word — the LoRA is always on once loaded, so strength is the only control. [source: model metadata]", + "Author never states which generation mode it was trained for; test against your own workflow before relying on it. [source: description]" + ], + "file": { + "name": "vhs_krea2.safetensors", + "size_bytes": 114324720, + "size_mb": 109.03, + "sha256": "cef79e53f88941e29ea7211543f624981207e32f15abef80c56fea3d0676e533" + }, + "stats": { + "downloads": 30523, + "thumbs_up": 2334, + "thumbs_down": 2 + }, + "popularity": { + "downloads": 30523, + "thumbs_up": 2334, + "thumbs_down": 2, + "comments": 0, + "approval": "100% positive (2334/2336)", + "age_days": 60.5, + "downloads_per_day": 89.92, + "rank_by_downloads": 18, + "rank_by_momentum": 30, + "of_total": 150, + "adoption_tier": "High", + "momentum": "Trending" + }, + "selection_score": 58531.0, + "selection_reasons": [ + "30,523 downloads", + "2,334 thumbs-up", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "1472x1024", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "1472x1024", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "1472x1024", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "1024x1472", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "1408x1088", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "1408x1088", + "prompt": "" + } + ], + "tags": [ + "vhs", + "2000s", + "crt", + "old", + "90s", + "tape", + "style", + "vintage" + ], + "scraped_at": "2026-09-05T00:11:54+00:00", + "path": "krea-2/2000s-analog-core" +} diff --git a/2000s-analog-core/vhs_krea2.safetensors b/2000s-analog-core/vhs_krea2.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..47d1803038ed73e76ed049c10e842e9cacd74444 --- /dev/null +++ b/2000s-analog-core/vhs_krea2.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cef79e53f88941e29ea7211543f624981207e32f15abef80c56fea3d0676e533 +size 114324720 diff --git a/aesthetic-amateur-photo/Krea_Amateur_V4.safetensors b/aesthetic-amateur-photo/Krea_Amateur_V4.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..065a1f8ba4d8f29f75ee0d4f35e295f43938f75f --- /dev/null +++ b/aesthetic-amateur-photo/Krea_Amateur_V4.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ad3b5092c99b4a1696df4ab07407775309da6c21d36c9a506f6b5ff153552ec +size 228588912 diff --git a/aesthetic-amateur-photo/example_images/01_image.jpeg b/aesthetic-amateur-photo/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..88013d6af2f2c84eb1265ae2cefa09902c95e275 --- /dev/null +++ b/aesthetic-amateur-photo/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd0e4fb03237679c6ce1479778d1d69db4fec090ad36bf11853151126a0df7df +size 2301518 diff --git a/aesthetic-amateur-photo/example_images/02_image.jpeg b/aesthetic-amateur-photo/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..3358383144a6680616d09495bff7165d72efb694 --- /dev/null +++ b/aesthetic-amateur-photo/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:60ca881cd29905cf83dba1b239658b5c050b69e93ae57f3f747a85783e319212 +size 2180862 diff --git a/aesthetic-amateur-photo/example_images/03_image.jpeg b/aesthetic-amateur-photo/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..a64351c5d7746d21afb63d6b9dcf371801202561 --- /dev/null +++ b/aesthetic-amateur-photo/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:edd8d4cd7f59b02a181fc56efba9e49a0ccdbab907cbd498b49ce3f16dfd224a +size 1968264 diff --git a/aesthetic-amateur-photo/example_images/04_image.jpeg b/aesthetic-amateur-photo/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..1b6b045d990be5e3a94fe03c48f3f8cd20ec2b18 --- /dev/null +++ b/aesthetic-amateur-photo/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bd90d43f48f77f541efb01f366ab98c518362ca9006f5dc74f297ed8ed2f26c4 +size 2834677 diff --git a/aesthetic-amateur-photo/example_images/05_image.jpeg b/aesthetic-amateur-photo/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..cf97667340785919e71ed188e493fa38043cfd0f --- /dev/null +++ b/aesthetic-amateur-photo/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aee8095af9e8d2083b2def71e0a7c246620e68d4e6225c13370af6c080337f4b +size 2328432 diff --git a/aesthetic-amateur-photo/example_images/06_image.jpeg b/aesthetic-amateur-photo/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..f1870ad962ab054f3b00d6337fc676ab25065454 --- /dev/null +++ b/aesthetic-amateur-photo/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7adf5a9e94e1c4c4ee0c34c7af537b726dfce2733c8b969d7dd62ab41798559d +size 2425696 diff --git a/aesthetic-amateur-photo/info.txt b/aesthetic-amateur-photo/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..a50c5076bcd81fdd3b336f2bcb48a2a39f714408 --- /dev/null +++ b/aesthetic-amateur-photo/info.txt @@ -0,0 +1,333 @@ +═══════════════════════════════════════════════════════════════════════════════ + AESTHETIC AMATEUR PHOTO +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Photorealism + Author powerpuff + Source https://civitai.com/models/689192 + Provenance CivitAI · model 689192 · version 3284331 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File Krea_Amateur_V4.safetensors + Version Krea V4 + Size 218.0 MB + Format SafeTensor / unknown precision + SHA256 0ad3b5092c99b4a1696df4ab07407775309da6c21d36c9a506f6b5ff153552ec + Published 2026-08-31 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + No trigger word. The LoRA applies as soon as it is loaded — + strength is the only way to dial the effect up or down. + +─────────────────────────────────────────────────────────────────────────────── + RECOMMENDED SETTINGS +─────────────────────────────────────────────────────────────────────────────── + The author did not publish specific settings. + No author settings were published. Community baseline: steps Turbo: 8 + (official README and HF card CLI example; the ComfyUI template ships 8). + Raw: 52 (official README). The krea-2 CLI's generic default is 28. + Community Turbo LoRAs commonly suggest 8-12 (e.g. the Identity Edit card + recommends Turbo at 8-12, or Raw at 20 for removals). A community 4-step + LoRA on Turbo makes 4 valid; its author calls 2 steps an out-of-spec + preview mode., cfg Turbo: CFG disabled - guidance_scale 0.0 in diffusers + and --cfg 0.0 in the krea-2 CLI, which is cfg 1.0 in ComfyUI's KSampler + (the official template ships cfg 1). These are the same thing; do not set + ComfyUI cfg to 0. Raw: CFG 3.5 (official README; CLI generic default 4.5). + Negative prompting on Turbo is effectively unused - the official templates + route ConditioningZeroOut into the negative slot., sampler ComfyUI: + KSampler with sampler euler, scheduler simple, denoise 1.0 (verified in + image_krea2_turbo_t2i.json: [seed, 'randomize', 8, 1, 'euler', 'simple', + 1]). Timestep shift mu = 1.15 for Turbo; ComfyUI bakes this in via + Krea2.sampling_settings = {"multiplier": 1.0, "shift": 1.15}, so the plain + t2i template needs no ModelSampling node, while the style-reference + template adds ModelSamplingFlux at max_shift 1.15 / base_shift 0.5. The CLI + exposes --mu 1.15 for Turbo; Raw uses a resolution-derived mu interpolated + between --y1 0.5 and --y2 1.15 (ai-toolkit mirrors this as base_shift 0.5 / + max_shift 1.15 over image_seq_len 256-6400, exponential time shift)., + resolution 1024x1024 default. Turbo: 1K-2K (the official CLI example + generates 2048x2048; ComfyUI's ResolutionSelector defaults to 1 megapixel, + set 2.0 for 2K). Raw: "trained to generate upto 1k resolution". + Width/height are padded up to a multiple of 16., lora_strength 1.0 for + Krea's nine official style LoRAs (docs.comfy.org lists Recommended Strength + 1.0 for every one), though the shipped t2i template's LoraLoaderModelOnly + is set to 0.8. The style-reference LoRA is used at 1.0 in its template. The + 4-step distill LoRA documents 1.0 as its trained point, 1.0-1.5 as coherent + extrapolation and above ~1.5 as speckle artefacting. Community subject + LoRAs are typically recommended around 0.8-1.2 and dropped toward 0.6-0.9 + when several are stacked.. + +─────────────────────────────────────────────────────────────────────────────── + ADOPTION & POPULARITY +─────────────────────────────────────────────────────────────────────────────── + Downloads 32,058 + Rating 100% positive (1413/1415) + Comments 0 + Published 4 days ago + Download rate 326.4/day since release + + Adoption rank #16 of 150 in this collection (High) + Momentum rank #5 of 150 by download rate (Hot · fresh release) + + Adoption tier is the percentile of total downloads within this collection; + momentum is the percentile of downloads-per-day since release. Momentum is + ranked rather than measured against a fixed rate, because this base model + is itself new and every LoRA looks fast on an absolute scale. Both numbers + are CivitAI's, read on the scrape date at the foot of this file. + +─────────────────────────────────────────────────────────────────────────────── + CHAINING / STACKING +─────────────────────────────────────────────────────────────────────────────── + Stacks with others not stated + + The author gave no chaining guidance. The rules below are the + general Krea 2 ones, not anything specific to this LoRA. + + General Krea 2 rules that apply here (see REFERENCE.txt): + + • The most important rule is not about chaining but about which + checkpoint you chain onto: the official README says in bold "TRAIN on + Raw and RUN on Turbo", and "We highly recommend using RAW for training + LoRAs and applying them on Turbo for inference." LoRAs trained on Raw + are stated to work well on Turbo. + • Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 + LoRAs are DiT-only, so there is no CLIP side to chain and the text- + encoder branch stays untouched however many LoRAs you add. + • Krea 2 LoRAs stack in practice and the community treats multi-LoRA as + routine - style + character, style + slider, realism + character. + Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one + writes "I recently stacked over 20 of my own LoRAs on a single + generation and the output still came out great." + • Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge + on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is + loaded. The CivitAI 12GB training guide also notes rank affects + composability: "rank 8 produced a lighter style influence that was + easier to stack with other LoRAs, rank 16 provided a useful balance + between concept strength and composability, and rank 32 was better + suited to a precise object or concept intended to take priority over + other stacked LoRAs." + • Accelerator LoRAs are position-sensitive in intent, not in node order: + a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a + sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to- + Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is + explicitly recommended by the 4-step LoRA's author. + • Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI + (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card + says "Keep mu = 1.15. The training targets are anchored to that grid; a + different shift moves the [sampler off them]"; the same reasoning + applies to any Turbo-trained adapter in the chain. + • Reference/edit LoRAs are designed to be stacked under ordinary subject + and style LoRAs - the Identity Edit card states "Composes with your + LoRAs: character/body/style LoRAs stack on top and steer the prior" and + recommends a stacked subject LoRA when unusual subjects drift toward + the base prior. + • Krea 2 LoRAs cannot be chained with LoRAs for any other base model. + Many CivitAI pages carry multi-model titles like "(Krea-2 + ZIT)" or + "[Flux | ZIB | Krea2]" - those are separate files for separate + architectures shipped under one model page, not a combined adapter. + • To combine two full Krea 2 checkpoints rather than adapters, ComfyUI + ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per- + section weights for first., tmlp., txtmlp., tproj., + txtfusion.projector, txtfusion.layerwise_blocks.0-1, + txtfusion.refiner_blocks.0-1, blocks.0-27 and last. + • The author never states which weight family this targets. Silent + degradation, never a hard error, between Raw and Turbo. The two share + one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" + - so every LoRA tensor shape matches on both and ComfyUI/diffuser + +─────────────────────────────────────────────────────────────────────────────── + PROS +─────────────────────────────────────────────────────────────────────────────── + + Widely adopted — 32,058 downloads. [source: CivitAI stats] + + Positively received — 1,413 thumbs-up, 2 down. [source: CivitAI stats] + + Compact at 218 MB — cheap to stack with other LoRAs. [source: file + size] + +─────────────────────────────────────────────────────────────────────────────── + CONS & CAVEATS +─────────────────────────────────────────────────────────────────────────────── + - Author labels this experimental — results vary shot to shot. [source: + description] + - No trigger word — the LoRA is always on once loaded, so strength is the + only control. [source: model metadata] + - Author never states which generation mode it was trained for; test + against your own workflow before relying on it. [source: description] + +─────────────────────────────────────────────────────────────────────────────── + COMPATIBILITY WARNING +─────────────────────────────────────────────────────────────────────────────── + Silent degradation, never a hard error, between Raw and Turbo. The two + share one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" - so + every LoRA tensor shape matches on both and ComfyUI/diffusers will load an + adapter against either without complaint. The lvladikov 4-step LoRA card + states this for the reverse direction: "Every layer it targets also exists + in Krea 2 Raw, so it will load there without complaint - but that is a side + effect of the shared architecture, not a supported mode", and reports + mixed, subject-dependent results with malformed subjects (duplicated heads, + fused limbs) at low step counts, needing 14+ steps with Raw's CFG on before + output is coherent. The intended direction (train on Raw, run on Turbo) is + officially blessed. The more common practical failure is a settings + mismatch rather than a weight mismatch: running a Turbo-targeted LoRA at + Raw's 52 steps + CFG 3.5, or leaving CFG enabled on Turbo, degrades output + with no warning. Loading a Krea 2 LoRA against a different model family + entirely (Flux, Qwen-Image, Z-Image) is also not an error - ComfyUI logs + "lora key not loaded: {key}" per key and generates as if no LoRA were + applied. + +─────────────────────────────────────────────────────────────────────────────── + AUTHOR'S DESCRIPTION +─────────────────────────────────────────────────────────────────────────────── + Another amateur photo LoRA aimed at making images feel more realistic and + closer to everyday smartphone or social media photos. Expect sharper + details, brighter colors, and a more casual, natural look overall. + + Krea V4: + Higher quality dataset, more steps, better captions. + + Krea V3: + Similar to V2 but the captions used a tag-style instead of full sentences. + + Krea V2: + More steps, more images, completely new captions + + Krea V1: + Initial release trained with basic captions. + + Z-Image Exp 1: + Experimental version trained with chinese captions. + Only 2000 steps and a small dataset but produces pretty good images. + Trigger: + 这张照片是用iPhone 13拍摄的。 + + Z-Image V2: + More than doubled the steps compared to V1 + Stronger effect + Can be used with higher strength + + Z-Image V1: + A new version for Z-Image trained on a small dataset. + + V4 Beta 2: + Trained on more images. + + V4 Beta 1: + I have started to create a completely new dataset with only high quality + images. For Beta 1 I used 111 images, though I plan to significantly + increase it for Beta 2. All images only have the caption "43stet1c", but it + isn't needed to mention it in your prompt. + This combination seems to produce the best and most consistent results so + far. + Increase the strength if the effect isn't strong enough. + + V3: + Same dataset as V2 but completely recaptioned with Claude 3.5 Sonnet and + upscaled to 1024 with Gigapixel. + + V2: + Trained on 124 images with no captions at all. Let me know if this works + better than V1. Results can be inconsistent, use HiRes fix and Beta + scheduler. + + V1: + This was trained on a dataset of 61 "aesthetic" images captioned with + Claude 3.5 Sonnet. + +─────────────────────────────────────────────────────────────────────────────── + VERSION NOTES +─────────────────────────────────────────────────────────────────────────────── + Higher quality dataset, more steps, better captions. + +─────────────────────────────────────────────────────────────────────────────── + EXAMPLE MEDIA +─────────────────────────────────────────────────────────────────────────────── + 6 preview(s) in ./example_images/ — community results for + this LoRA. Videos keep a matching _poster.jpg still frame. + + No prompts are recorded below: CivitAI's public API flags these posts as + having generation metadata but does not return it, so the prompts cannot be + scraped. Open the model page to read them. + + 01_image.jpeg (image, 1024x1536) + 02_image.jpeg (image, 1088x1472) + 03_image.jpeg (image, 1024x1536) + 04_image.jpeg (image, 1088x1472) + 05_image.jpeg (image, 1024x1536) + 06_image.jpeg (image, 1024x1536) + +─────────────────────────────────────────────────────────────────────────────── + HOW TO USE (COMFYUI) +─────────────────────────────────────────────────────────────────────────────── + 1. Update ComfyUI to at least v0.26.0 (released 2026-06-23), which added + local Krea2 support. For reference-image / style-reference / edit LoRAs you + need at least v0.29.0 (2026-07-29), which added regular and timestep-zero + reference images for Krea 2. ModelMergeKrea2 arrived in v0.27.0 + (2026-06-30). + + 2. Put the base stack in place (files from huggingface.co/Comfy-Org/Krea-2, + which is not gated): + - ComfyUI/models/diffusion_models/krea2_turbo_fp8_scaled.safetensors + (documented recommended default; the repo also holds krea2_turbo_bf16, + _int8_convrot, _mxfp8, _nvfp4 and krea2_raw_bf16 / _fp8_scaled / + _int8_convrot) + - ComfyUI/models/text_encoders/qwen3vl_4b_fp8_scaled.safetensors (bf16 + also available) + - ComfyUI/models/vae/qwen_image_vae.safetensors + + 3. Put the LoRA .safetensors in ComfyUI/models/loras/ - flat, no per-model + subfolder required. + + 4. Build or load the graph. The official image_krea2_turbo_t2i.json + template is: UNETLoader (krea2_turbo_fp8_scaled.safetensors) -> + LoraLoaderModelOnly -> KSampler; CLIPLoader + (qwen3vl_4b_fp8_scaled.safetensors, type "krea2") -> text encode -> + KSampler positive, with ConditioningZeroOut into negative; VAELoader + (qwen_image_vae.safetensors) -> VAEDecode -> SaveImage; EmptyLatentImage + 1024x1024 driven by ResolutionSelector. An int8 sibling template + (image_krea2_turbo_t2i_int8.json) also ships. + + 5. Load the LoRA with LoraLoaderModelOnly, not LoraLoader - both official + templates do. Krea 2 LoRAs patch the DiT only; note that the DiT's own + text_fusion / txtfusion blocks are a normal LoRA target (the official style + LoRAs patch them), but the Qwen3-VL text encoder itself is not. Set + strength 1.0 for Krea's official style LoRAs (the shipped template uses + 0.8) and add the LoRA's trigger word to the prompt. + + 6. Sampler settings: steps 8, cfg 1, sampler euler, scheduler simple, + denoise 1.0. Do not add a ModelSampling node for plain t2i - ComfyUI + already applies shift 1.15 from Krea2.sampling_settings. + + 7. For style reference, use the dedicated template: UNETLoader + krea2_turbo_int8_convrot.safetensors, LoraLoaderModelOnly + krea2_style_reference.safetensors at 1.0, ModelSamplingFlux (1.15, 0.5), + TextEncodeQwenImageEditPlus fed by LoadImage, + FluxKontextMultiReferenceLatentMethod set to index_timestep_zero, and + SamplerCustomAdvanced with KSamplerSelect euler + BasicScheduler simple/8 + + CFGGuider cfg 1. + + 8. For instruction editing (Krea 2 Identity Edit), stock nodes are not + enough: install the ComfyUI-Krea2Edit node pack, which ships its own + workflows, and run Turbo at 8-12 steps / cfg 1 (or Raw at 20 steps / cfg 3 + for removals) with LoRA strength 1.0. + + 9. Low-VRAM alternative: install city96's ComfyUI-GGUF and use a community + GGUF build (realrebelai/KREA-2_GGUFs, quantized from Turbo; + molbal/krea2-gguf, quantized from Raw) with the Unet Loader (GGUF) node in + place of UNETLoader. There is no official GGUF release. LoRA loading is + unchanged. + +─────────────────────────────────────────────────────────────────────────────── + Scraped by lorakit on 2026-09-05 00:12 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/aesthetic-amateur-photo/metadata.json b/aesthetic-amateur-photo/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..5f1c7e8ca6fa7acc40cf5ef554d71c7f1085e122 --- /dev/null +++ b/aesthetic-amateur-photo/metadata.json @@ -0,0 +1,131 @@ +{ + "name": "Aesthetic Amateur Photo", + "slug": "aesthetic-amateur-photo", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 689192, + "version_id": 3284331, + "version_name": "Krea V4", + "source": "civitai", + "url": "https://civitai.com/models/689192", + "repo_id": null, + "file_path": null, + "author": "powerpuff", + "category": { + "key": "photoreal", + "label": "Photorealism" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [], + "recommended_settings": {}, + "chaining": { + "author": { + "stacks": "not stated", + "notes": "", + "avoid": [], + "requires": [], + "order": "", + "strength": "", + "stated": false + }, + "general_rules": [ + "The most important rule is not about chaining but about which checkpoint you chain onto: the official README says in bold \"TRAIN on Raw and RUN on Turbo\", and \"We highly recommend using RAW for training LoRAs and applying them on Turbo for inference.\" LoRAs trained on Raw are stated to work well on Turbo.", + "Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 LoRAs are DiT-only, so there is no CLIP side to chain and the text-encoder branch stays untouched however many LoRAs you add.", + "Krea 2 LoRAs stack in practice and the community treats multi-LoRA as routine - style + character, style + slider, realism + character. Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one writes \"I recently stacked over 20 of my own LoRAs on a single generation and the output still came out great.\"", + "Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is loaded. The CivitAI 12GB training guide also notes rank affects composability: \"rank 8 produced a lighter style influence that was easier to stack with other LoRAs, rank 16 provided a useful balance between concept strength and composability, and rank 32 was better suited to a precise object or concept intended to take priority over other stacked LoRAs.\"", + "Accelerator LoRAs are position-sensitive in intent, not in node order: a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to-Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is explicitly recommended by the 4-step LoRA's author.", + "Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card says \"Keep mu = 1.15. The training targets are anchored to that grid; a different shift moves the [sampler off them]\"; the same reasoning applies to any Turbo-trained adapter in the chain.", + "Reference/edit LoRAs are designed to be stacked under ordinary subject and style LoRAs - the Identity Edit card states \"Composes with your LoRAs: character/body/style LoRAs stack on top and steer the prior\" and recommends a stacked subject LoRA when unusual subjects drift toward the base prior.", + "Krea 2 LoRAs cannot be chained with LoRAs for any other base model. Many CivitAI pages carry multi-model titles like \"(Krea-2 + ZIT)\" or \"[Flux | ZIB | Krea2]\" - those are separate files for separate architectures shipped under one model page, not a combined adapter.", + "To combine two full Krea 2 checkpoints rather than adapters, ComfyUI ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per-section weights for first., tmlp., txtmlp., tproj., txtfusion.projector, txtfusion.layerwise_blocks.0-1, txtfusion.refiner_blocks.0-1, blocks.0-27 and last.", + "The author never states which weight family this targets. Silent degradation, never a hard error, between Raw and Turbo. The two share one architecture - ai-toolkit's config is commented \"The reference 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)\" - so every LoRA tensor shape matches on both and ComfyUI/diffuser" + ] + }, + "pros": [ + "Widely adopted — 32,058 downloads. [source: CivitAI stats]", + "Positively received — 1,413 thumbs-up, 2 down. [source: CivitAI stats]", + "Compact at 218 MB — cheap to stack with other LoRAs. [source: file size]" + ], + "cons": [ + "Author labels this experimental — results vary shot to shot. [source: description]", + "No trigger word — the LoRA is always on once loaded, so strength is the only control. [source: model metadata]", + "Author never states which generation mode it was trained for; test against your own workflow before relying on it. [source: description]" + ], + "file": { + "name": "Krea_Amateur_V4.safetensors", + "size_bytes": 228588912, + "size_mb": 218.0, + "sha256": "0ad3b5092c99b4a1696df4ab07407775309da6c21d36c9a506f6b5ff153552ec" + }, + "stats": { + "downloads": 32058, + "thumbs_up": 1413, + "thumbs_down": 2 + }, + "popularity": { + "downloads": 32058, + "thumbs_up": 1413, + "thumbs_down": 2, + "comments": 0, + "approval": "100% positive (1413/1415)", + "age_days": 4.4, + "downloads_per_day": 326.41, + "rank_by_downloads": 16, + "rank_by_momentum": 5, + "of_total": 150, + "adoption_tier": "High", + "momentum": "Hot · fresh release" + }, + "selection_score": 49014.0, + "selection_reasons": [ + "32,058 downloads", + "1,413 thumbs-up", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "1024x1536", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "1088x1472", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "1024x1536", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "1088x1472", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "1024x1536", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "1024x1536", + "prompt": "" + } + ], + "tags": [ + "photorealistic", + "style" + ], + "scraped_at": "2026-09-05T00:12:07+00:00", + "path": "krea-2/aesthetic-amateur-photo" +} diff --git a/background-detail-enhancer/example_images/01_image.jpeg b/background-detail-enhancer/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..e83376aaa6a70e9aff54c062e41d35f78dbc0aba --- /dev/null +++ b/background-detail-enhancer/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ad9f6f1f191fe7629616cb0e8de5c26a328a02107a11705846e7af7d4506b54d +size 1469951 diff --git a/background-detail-enhancer/example_images/02_image.jpeg b/background-detail-enhancer/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..c2d3a5355f0461d041bc634b68d0e3e712158446 --- /dev/null +++ b/background-detail-enhancer/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:31a79be70258cd9a964924557d0556f069b08e45127c8e6477dcdf756f349c42 +size 1353742 diff --git a/background-detail-enhancer/example_images/03_image.jpeg b/background-detail-enhancer/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..85f3aca9fb025f5c053a944ea21240ae8412f3e1 --- /dev/null +++ b/background-detail-enhancer/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a115046f92a129ff2f40020aa354844e91ebd18d902ebaaa25e97f6f6bd66b9 +size 538878 diff --git a/background-detail-enhancer/example_images/04_image.jpeg b/background-detail-enhancer/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..faa148cf09c5a13a712e507adc6f90151c41e1af --- /dev/null +++ b/background-detail-enhancer/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2752d4f9dc948fd3d0b13d81533e607aba263444a8781e07727ebc26072c0323 +size 1407216 diff --git a/background-detail-enhancer/example_images/05_image.jpeg b/background-detail-enhancer/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..d786c102cd99131c68ad93eb96c905e0eb9c9483 --- /dev/null +++ b/background-detail-enhancer/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57cf9f5515003e14e7bee2a5d7f2c6dacb41876137bf0acb0944b860efb5bad1 +size 1425417 diff --git a/background-detail-enhancer/example_images/06_image.jpeg b/background-detail-enhancer/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..c4be8dd785b772e4162668b01d46b6b23f016db5 --- /dev/null +++ b/background-detail-enhancer/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:862fa9dcd9262a0cb51ca9228cd9646584774fc3c83be6ebb82adcc6893d2943 +size 1585467 diff --git a/background-detail-enhancer/info.txt b/background-detail-enhancer/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..2937ae3c5712737d4cc5c28e7212ae6adb40bd16 --- /dev/null +++ b/background-detail-enhancer/info.txt @@ -0,0 +1,277 @@ +═══════════════════════════════════════════════════════════════════════════════ + BACKGROUND DETAIL ENHANCER✨ +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Detail / anatomy fixer + Author ZyloO + Source https://civitai.com/models/633524 + Provenance CivitAI · model 633524 · version 3211112 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File zy_BackgroundDetail_K2.safetensors + Version Krea 2 + Size 218.0 MB + Format SafeTensor / unknown precision + SHA256 190a8835e96bcae59853cf7025ee0dad3a5963f240b945cb26e03798d12ccbba + Published 2026-08-09 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + No trigger word. The LoRA applies as soon as it is loaded — + strength is the only way to dial the effect up or down. + +─────────────────────────────────────────────────────────────────────────────── + RECOMMENDED SETTINGS +─────────────────────────────────────────────────────────────────────────────── + Taken verbatim from the author's description: + + LoRA strength 0.1–0.75 + +─────────────────────────────────────────────────────────────────────────────── + ADOPTION & POPULARITY +─────────────────────────────────────────────────────────────────────────────── + Downloads 23,711 + Rating 100% positive (2830/2831) + Comments 12 + Published 26 days ago + Download rate 70.2/day since release + + Adoption rank #24 of 150 in this collection (High) + Momentum rank #39 of 150 by download rate (Steady · fresh release) + + Adoption tier is the percentile of total downloads within this collection; + momentum is the percentile of downloads-per-day since release. Momentum is + ranked rather than measured against a fixed rate, because this base model + is itself new and every LoRA looks fast on an absolute scale. Both numbers + are CivitAI's, read on the scrape date at the foot of this file. + +─────────────────────────────────────────────────────────────────────────────── + CHAINING / STACKING +─────────────────────────────────────────────────────────────────────────────── + Stacks with others not stated + + The author gave no chaining guidance. The rules below are the + general Krea 2 ones, not anything specific to this LoRA. + + General Krea 2 rules that apply here (see REFERENCE.txt): + + • The most important rule is not about chaining but about which + checkpoint you chain onto: the official README says in bold "TRAIN on + Raw and RUN on Turbo", and "We highly recommend using RAW for training + LoRAs and applying them on Turbo for inference." LoRAs trained on Raw + are stated to work well on Turbo. + • Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 + LoRAs are DiT-only, so there is no CLIP side to chain and the text- + encoder branch stays untouched however many LoRAs you add. + • Krea 2 LoRAs stack in practice and the community treats multi-LoRA as + routine - style + character, style + slider, realism + character. + Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one + writes "I recently stacked over 20 of my own LoRAs on a single + generation and the output still came out great." + • Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge + on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is + loaded. The CivitAI 12GB training guide also notes rank affects + composability: "rank 8 produced a lighter style influence that was + easier to stack with other LoRAs, rank 16 provided a useful balance + between concept strength and composability, and rank 32 was better + suited to a precise object or concept intended to take priority over + other stacked LoRAs." + • Accelerator LoRAs are position-sensitive in intent, not in node order: + a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a + sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to- + Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is + explicitly recommended by the 4-step LoRA's author. + • Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI + (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card + says "Keep mu = 1.15. The training targets are anchored to that grid; a + different shift moves the [sampler off them]"; the same reasoning + applies to any Turbo-trained adapter in the chain. + • Reference/edit LoRAs are designed to be stacked under ordinary subject + and style LoRAs - the Identity Edit card states "Composes with your + LoRAs: character/body/style LoRAs stack on top and steer the prior" and + recommends a stacked subject LoRA when unusual subjects drift toward + the base prior. + • Krea 2 LoRAs cannot be chained with LoRAs for any other base model. + Many CivitAI pages carry multi-model titles like "(Krea-2 + ZIT)" or + "[Flux | ZIB | Krea2]" - those are separate files for separate + architectures shipped under one model page, not a combined adapter. + • To combine two full Krea 2 checkpoints rather than adapters, ComfyUI + ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per- + section weights for first., tmlp., txtmlp., tproj., + txtfusion.projector, txtfusion.layerwise_blocks.0-1, + txtfusion.refiner_blocks.0-1, blocks.0-27 and last. + • The author never states which weight family this targets. Silent + degradation, never a hard error, between Raw and Turbo. The two share + one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" + - so every LoRA tensor shape matches on both and ComfyUI/diffuser + +─────────────────────────────────────────────────────────────────────────────── + PROS +─────────────────────────────────────────────────────────────────────────────── + + Widely adopted — 23,711 downloads. [source: CivitAI stats] + + Positively received — 2,830 thumbs-up, 1 down. [source: CivitAI stats] + + Compact at 218 MB — cheap to stack with other LoRAs. [source: file + size] + + Author published concrete recommended settings. [source: description] + +─────────────────────────────────────────────────────────────────────────────── + CONS & CAVEATS +─────────────────────────────────────────────────────────────────────────────── + - No trigger word — the LoRA is always on once loaded, so strength is the + only control. [source: model metadata] + - Author never states which generation mode it was trained for; test + against your own workflow before relying on it. [source: description] + +─────────────────────────────────────────────────────────────────────────────── + COMPATIBILITY WARNING +─────────────────────────────────────────────────────────────────────────────── + Silent degradation, never a hard error, between Raw and Turbo. The two + share one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" - so + every LoRA tensor shape matches on both and ComfyUI/diffusers will load an + adapter against either without complaint. The lvladikov 4-step LoRA card + states this for the reverse direction: "Every layer it targets also exists + in Krea 2 Raw, so it will load there without complaint - but that is a side + effect of the shared architecture, not a supported mode", and reports + mixed, subject-dependent results with malformed subjects (duplicated heads, + fused limbs) at low step counts, needing 14+ steps with Raw's CFG on before + output is coherent. The intended direction (train on Raw, run on Turbo) is + officially blessed. The more common practical failure is a settings + mismatch rather than a weight mismatch: running a Turbo-targeted LoRA at + Raw's 52 steps + CFG 3.5, or leaving CFG enabled on Turbo, degrades output + with no warning. Loading a Krea 2 LoRA against a different model family + entirely (Flux, Qwen-Image, Z-Image) is also not an error - ComfyUI logs + "lora key not loaded: {key}" per key and generates as if no LoRA were + applied. + +─────────────────────────────────────────────────────────────────────────────── + AUTHOR'S DESCRIPTION +─────────────────────────────────────────────────────────────────────────────── + ✨ REFERRAL CODE : DMD9XGMN ℹ️ Model Overview + - This LoRA enhances the background of the image, adding more detail and + clutter to it. + + - Image Comparison + + 🏠 Recommended Parameters Pony Settings: - Strength : 0.1-0.75 . + + - Clip Skip : 2 + + Additional Information: - Useful Prompt Words: + - Clutter + + - Background + + - Detailed background + + - Elaborate + + - Rich + + - Dynamic + + - Immersive + + - Intricate + + 🏋️‍♀️ Training LoRas - Read the following article for tips and my training + preset + + 🌞 Requests - Do you want a specific LoRA? You can request a commission on + my Ko-Fi + + ⚡️ Buzz for the Best Images ⚡️ + +─────────────────────────────────────────────────────────────────────────────── + EXAMPLE MEDIA +─────────────────────────────────────────────────────────────────────────────── + 6 preview(s) in ./example_images/ — community results for + this LoRA. Videos keep a matching _poster.jpg still frame. + + No prompts are recorded below: CivitAI's public API flags these posts as + having generation metadata but does not return it, so the prompts cannot be + scraped. Open the model page to read them. + + 01_image.jpeg (image, 768x1024) + 02_image.jpeg (image, 768x1024) + 03_image.jpeg (image, 768x1376) + 04_image.jpeg (image, 768x1024) + 05_image.jpeg (image, 768x1024) + 06_image.jpeg (image, 768x1024) + +─────────────────────────────────────────────────────────────────────────────── + HOW TO USE (COMFYUI) +─────────────────────────────────────────────────────────────────────────────── + 1. Update ComfyUI to at least v0.26.0 (released 2026-06-23), which added + local Krea2 support. For reference-image / style-reference / edit LoRAs you + need at least v0.29.0 (2026-07-29), which added regular and timestep-zero + reference images for Krea 2. ModelMergeKrea2 arrived in v0.27.0 + (2026-06-30). + + 2. Put the base stack in place (files from huggingface.co/Comfy-Org/Krea-2, + which is not gated): + - ComfyUI/models/diffusion_models/krea2_turbo_fp8_scaled.safetensors + (documented recommended default; the repo also holds krea2_turbo_bf16, + _int8_convrot, _mxfp8, _nvfp4 and krea2_raw_bf16 / _fp8_scaled / + _int8_convrot) + - ComfyUI/models/text_encoders/qwen3vl_4b_fp8_scaled.safetensors (bf16 + also available) + - ComfyUI/models/vae/qwen_image_vae.safetensors + + 3. Put the LoRA .safetensors in ComfyUI/models/loras/ - flat, no per-model + subfolder required. + + 4. Build or load the graph. The official image_krea2_turbo_t2i.json + template is: UNETLoader (krea2_turbo_fp8_scaled.safetensors) -> + LoraLoaderModelOnly -> KSampler; CLIPLoader + (qwen3vl_4b_fp8_scaled.safetensors, type "krea2") -> text encode -> + KSampler positive, with ConditioningZeroOut into negative; VAELoader + (qwen_image_vae.safetensors) -> VAEDecode -> SaveImage; EmptyLatentImage + 1024x1024 driven by ResolutionSelector. An int8 sibling template + (image_krea2_turbo_t2i_int8.json) also ships. + + 5. Load the LoRA with LoraLoaderModelOnly, not LoraLoader - both official + templates do. Krea 2 LoRAs patch the DiT only; note that the DiT's own + text_fusion / txtfusion blocks are a normal LoRA target (the official style + LoRAs patch them), but the Qwen3-VL text encoder itself is not. Set + strength 1.0 for Krea's official style LoRAs (the shipped template uses + 0.8) and add the LoRA's trigger word to the prompt. + + 6. Sampler settings: steps 8, cfg 1, sampler euler, scheduler simple, + denoise 1.0. Do not add a ModelSampling node for plain t2i - ComfyUI + already applies shift 1.15 from Krea2.sampling_settings. + + 7. For style reference, use the dedicated template: UNETLoader + krea2_turbo_int8_convrot.safetensors, LoraLoaderModelOnly + krea2_style_reference.safetensors at 1.0, ModelSamplingFlux (1.15, 0.5), + TextEncodeQwenImageEditPlus fed by LoadImage, + FluxKontextMultiReferenceLatentMethod set to index_timestep_zero, and + SamplerCustomAdvanced with KSamplerSelect euler + BasicScheduler simple/8 + + CFGGuider cfg 1. + + 8. For instruction editing (Krea 2 Identity Edit), stock nodes are not + enough: install the ComfyUI-Krea2Edit node pack, which ships its own + workflows, and run Turbo at 8-12 steps / cfg 1 (or Raw at 20 steps / cfg 3 + for removals) with LoRA strength 1.0. + + 9. Low-VRAM alternative: install city96's ComfyUI-GGUF and use a community + GGUF build (realrebelai/KREA-2_GGUFs, quantized from Turbo; + molbal/krea2-gguf, quantized from Raw) with the Unet Loader (GGUF) node in + place of UNETLoader. There is no official GGUF release. LoRA loading is + unchanged. + +─────────────────────────────────────────────────────────────────────────────── + Scraped by lorakit on 2026-09-05 00:12 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/background-detail-enhancer/metadata.json b/background-detail-enhancer/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..ff5cb9157da6835ddbc638f5dbaf4652bdfa6095 --- /dev/null +++ b/background-detail-enhancer/metadata.json @@ -0,0 +1,139 @@ +{ + "name": "Background Detail Enhancer✨", + "slug": "background-detail-enhancer", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 633524, + "version_id": 3211112, + "version_name": "Krea 2", + "source": "civitai", + "url": "https://civitai.com/models/633524", + "repo_id": null, + "file_path": null, + "author": "ZyloO", + "category": { + "key": "detail", + "label": "Detail / anatomy fixer" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [], + "recommended_settings": { + "LoRA strength": "0.1–0.75" + }, + "chaining": { + "author": { + "stacks": "not stated", + "notes": "", + "avoid": [], + "requires": [], + "order": "", + "strength": "", + "stated": false + }, + "general_rules": [ + "The most important rule is not about chaining but about which checkpoint you chain onto: the official README says in bold \"TRAIN on Raw and RUN on Turbo\", and \"We highly recommend using RAW for training LoRAs and applying them on Turbo for inference.\" LoRAs trained on Raw are stated to work well on Turbo.", + "Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 LoRAs are DiT-only, so there is no CLIP side to chain and the text-encoder branch stays untouched however many LoRAs you add.", + "Krea 2 LoRAs stack in practice and the community treats multi-LoRA as routine - style + character, style + slider, realism + character. Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one writes \"I recently stacked over 20 of my own LoRAs on a single generation and the output still came out great.\"", + "Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is loaded. The CivitAI 12GB training guide also notes rank affects composability: \"rank 8 produced a lighter style influence that was easier to stack with other LoRAs, rank 16 provided a useful balance between concept strength and composability, and rank 32 was better suited to a precise object or concept intended to take priority over other stacked LoRAs.\"", + "Accelerator LoRAs are position-sensitive in intent, not in node order: a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to-Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is explicitly recommended by the 4-step LoRA's author.", + "Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card says \"Keep mu = 1.15. The training targets are anchored to that grid; a different shift moves the [sampler off them]\"; the same reasoning applies to any Turbo-trained adapter in the chain.", + "Reference/edit LoRAs are designed to be stacked under ordinary subject and style LoRAs - the Identity Edit card states \"Composes with your LoRAs: character/body/style LoRAs stack on top and steer the prior\" and recommends a stacked subject LoRA when unusual subjects drift toward the base prior.", + "Krea 2 LoRAs cannot be chained with LoRAs for any other base model. Many CivitAI pages carry multi-model titles like \"(Krea-2 + ZIT)\" or \"[Flux | ZIB | Krea2]\" - those are separate files for separate architectures shipped under one model page, not a combined adapter.", + "To combine two full Krea 2 checkpoints rather than adapters, ComfyUI ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per-section weights for first., tmlp., txtmlp., tproj., txtfusion.projector, txtfusion.layerwise_blocks.0-1, txtfusion.refiner_blocks.0-1, blocks.0-27 and last.", + "The author never states which weight family this targets. Silent degradation, never a hard error, between Raw and Turbo. The two share one architecture - ai-toolkit's config is commented \"The reference 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)\" - so every LoRA tensor shape matches on both and ComfyUI/diffuser" + ] + }, + "pros": [ + "Widely adopted — 23,711 downloads. [source: CivitAI stats]", + "Positively received — 2,830 thumbs-up, 1 down. [source: CivitAI stats]", + "Compact at 218 MB — cheap to stack with other LoRAs. [source: file size]", + "Author published concrete recommended settings. [source: description]" + ], + "cons": [ + "No trigger word — the LoRA is always on once loaded, so strength is the only control. [source: model metadata]", + "Author never states which generation mode it was trained for; test against your own workflow before relying on it. [source: description]" + ], + "file": { + "name": "zy_BackgroundDetail_K2.safetensors", + "size_bytes": 228588904, + "size_mb": 218.0, + "sha256": "190a8835e96bcae59853cf7025ee0dad3a5963f240b945cb26e03798d12ccbba" + }, + "stats": { + "downloads": 23711, + "thumbs_up": 2830, + "thumbs_down": 1 + }, + "popularity": { + "downloads": 23711, + "thumbs_up": 2830, + "thumbs_down": 1, + "comments": 12, + "approval": "100% positive (2830/2831)", + "age_days": 26.5, + "downloads_per_day": 70.23, + "rank_by_downloads": 24, + "rank_by_momentum": 39, + "of_total": 150, + "adoption_tier": "High", + "momentum": "Steady · fresh release" + }, + "selection_score": 57671.0, + "selection_reasons": [ + "23,711 downloads", + "2,830 thumbs-up", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "768x1024", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "768x1024", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "768x1376", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "768x1024", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "768x1024", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "768x1024", + "prompt": "" + } + ], + "tags": [ + "scenery", + "backgrounds", + "enhancer", + "detailer", + "detail", + "concept", + "clutter", + "environment" + ], + "scraped_at": "2026-09-05T00:12:04+00:00", + "path": "krea-2/background-detail-enhancer" +} diff --git a/background-detail-enhancer/zy_BackgroundDetail_K2.safetensors b/background-detail-enhancer/zy_BackgroundDetail_K2.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..02cd0a9de574c605886a432e0aeb5847e951e258 --- /dev/null +++ b/background-detail-enhancer/zy_BackgroundDetail_K2.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:190a8835e96bcae59853cf7025ee0dad3a5963f240b945cb26e03798d12ccbba +size 228588904 diff --git a/bartstyle-blue-archive-art-style-lora-anime-flat-color/BArtstyle-KREA2-V2-1024.safetensors b/bartstyle-blue-archive-art-style-lora-anime-flat-color/BArtstyle-KREA2-V2-1024.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..c77931988364cb82734738c4522ee417f18892fc --- /dev/null +++ b/bartstyle-blue-archive-art-style-lora-anime-flat-color/BArtstyle-KREA2-V2-1024.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c19dc382d92002bfc53ece7df5556acc288579af9f0719752059dfae7fe9f70 +size 938539208 diff --git a/bartstyle-blue-archive-art-style-lora-anime-flat-color/example_images/01_image.jpeg b/bartstyle-blue-archive-art-style-lora-anime-flat-color/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..a776a92ff0f5e16204b4b1b707d56f5e6a4acb88 --- /dev/null +++ b/bartstyle-blue-archive-art-style-lora-anime-flat-color/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c7f579aa5b1324377c40f9cc32fafd835e6d4c706dbf4d9b83bc0dc763fcd842 +size 3583347 diff --git a/bartstyle-blue-archive-art-style-lora-anime-flat-color/example_images/02_image.jpeg b/bartstyle-blue-archive-art-style-lora-anime-flat-color/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..7b70aaf6e6956c8313c0932ba0b078b842a8fe57 --- /dev/null +++ b/bartstyle-blue-archive-art-style-lora-anime-flat-color/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:321de02051108238b9bede6551174d6e01da73043fec9339d58a1b8096e2485c +size 4369055 diff --git a/bartstyle-blue-archive-art-style-lora-anime-flat-color/example_images/03_image.jpeg b/bartstyle-blue-archive-art-style-lora-anime-flat-color/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..90a656eeba8c8001365fff520d8bdba683b21ea9 --- /dev/null +++ b/bartstyle-blue-archive-art-style-lora-anime-flat-color/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0690649526876aadb869a4a3455c62e0e2b90539adaee948c3918c4fc0b4d3a6 +size 3151329 diff --git a/bartstyle-blue-archive-art-style-lora-anime-flat-color/example_images/04_image.jpeg b/bartstyle-blue-archive-art-style-lora-anime-flat-color/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..a2167ab1d6cc28f7f3769e7cb72a0e38d1dec0c6 --- /dev/null +++ b/bartstyle-blue-archive-art-style-lora-anime-flat-color/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a367c3fdc547cc53e9ec8756a8c4b762305e580509dcc37abcbb422949504900 +size 153800 diff --git a/bartstyle-blue-archive-art-style-lora-anime-flat-color/example_images/05_image.jpeg b/bartstyle-blue-archive-art-style-lora-anime-flat-color/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..7101e97bf2791f778baf182e4a2c544cb00d6718 --- /dev/null +++ b/bartstyle-blue-archive-art-style-lora-anime-flat-color/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b4e69798974eb5742d4fb16e70596c066aebcecdd9867fec2a6489d4022715b7 +size 4526937 diff --git a/bartstyle-blue-archive-art-style-lora-anime-flat-color/example_images/06_image.jpeg b/bartstyle-blue-archive-art-style-lora-anime-flat-color/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..9ebf33d75dc3ab63831080907085f1a69f61e387 --- /dev/null +++ b/bartstyle-blue-archive-art-style-lora-anime-flat-color/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a11de186d351f80705115b0110b7c42638ba52fb1343d7f993ceb48301825e8d +size 3686998 diff --git a/bartstyle-blue-archive-art-style-lora-anime-flat-color/info.txt b/bartstyle-blue-archive-art-style-lora-anime-flat-color/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..3ba1dee9e6dae93882497ae97f7a36ba03ba9ff2 --- /dev/null +++ b/bartstyle-blue-archive-art-style-lora-anime-flat-color/info.txt @@ -0,0 +1,493 @@ +═══════════════════════════════════════════════════════════════════════════════ + BARTSTYLE | BLUE ARCHIVE ART STYLE LORA | ANIME FLAT COLOR | 蔚蓝档案画风模型 | + ブルーアーカイブ画風モデル +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Visual style + Author 4BEraser + Source https://civitai.com/models/29215 + Provenance CivitAI · model 29215 · version 3220090 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File BArtstyle-KREA2-V2-1024.safetensors + Version Krea2-V2 + Size 895.1 MB + Format SafeTensor / unknown precision + SHA256 9c19dc382d92002bfc53ece7df5556acc288579af9f0719752059dfae7fe9f70 + Published 2026-08-11 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + › "azure youth anime art style" + +─────────────────────────────────────────────────────────────────────────────── + RECOMMENDED SETTINGS +─────────────────────────────────────────────────────────────────────────────── + Taken verbatim from the author's description: + + LoRA strength 0.6–0.7 + Resolution 1024x768 + Sampler DPM++ 2M + +─────────────────────────────────────────────────────────────────────────────── + ADOPTION & POPULARITY +─────────────────────────────────────────────────────────────────────────────── + Downloads 28,534 + Rating 100% positive (2683/2684) + Comments 0 + Published 24 days ago + Download rate 82.3/day since release + + Adoption rank #21 of 150 in this collection (High) + Momentum rank #34 of 150 by download rate (Trending · fresh release) + + Adoption tier is the percentile of total downloads within this collection; + momentum is the percentile of downloads-per-day since release. Momentum is + ranked rather than measured against a fixed rate, because this base model + is itself new and every LoRA looks fast on an absolute scale. Both numbers + are CivitAI's, read on the scrape date at the foot of this file. + +─────────────────────────────────────────────────────────────────────────────── + CHAINING / STACKING +─────────────────────────────────────────────────────────────────────────────── + Stacks with others not stated + + The author gave no chaining guidance. The rules below are the + general Krea 2 ones, not anything specific to this LoRA. + + General Krea 2 rules that apply here (see REFERENCE.txt): + + • The most important rule is not about chaining but about which + checkpoint you chain onto: the official README says in bold "TRAIN on + Raw and RUN on Turbo", and "We highly recommend using RAW for training + LoRAs and applying them on Turbo for inference." LoRAs trained on Raw + are stated to work well on Turbo. + • Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 + LoRAs are DiT-only, so there is no CLIP side to chain and the text- + encoder branch stays untouched however many LoRAs you add. + • Krea 2 LoRAs stack in practice and the community treats multi-LoRA as + routine - style + character, style + slider, realism + character. + Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one + writes "I recently stacked over 20 of my own LoRAs on a single + generation and the output still came out great." + • Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge + on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is + loaded. The CivitAI 12GB training guide also notes rank affects + composability: "rank 8 produced a lighter style influence that was + easier to stack with other LoRAs, rank 16 provided a useful balance + between concept strength and composability, and rank 32 was better + suited to a precise object or concept intended to take priority over + other stacked LoRAs." + • Accelerator LoRAs are position-sensitive in intent, not in node order: + a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a + sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to- + Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is + explicitly recommended by the 4-step LoRA's author. + • Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI + (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card + says "Keep mu = 1.15. The training targets are anchored to that grid; a + different shift moves the [sampler off them]"; the same reasoning + applies to any Turbo-trained adapter in the chain. + • Reference/edit LoRAs are designed to be stacked under ordinary subject + and style LoRAs - the Identity Edit card states "Composes with your + LoRAs: character/body/style LoRAs stack on top and steer the prior" and + recommends a stacked subject LoRA when unusual subjects drift toward + the base prior. + • Krea 2 LoRAs cannot be chained with LoRAs for any other base model. + Many CivitAI pages carry multi-model titles like "(Krea-2 + ZIT)" or + "[Flux | ZIB | Krea2]" - those are separate files for separate + architectures shipped under one model page, not a combined adapter. + • To combine two full Krea 2 checkpoints rather than adapters, ComfyUI + ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per- + section weights for first., tmlp., txtmlp., tproj., + txtfusion.projector, txtfusion.layerwise_blocks.0-1, + txtfusion.refiner_blocks.0-1, blocks.0-27 and last. + • The author never states which weight family this targets. Silent + degradation, never a hard error, between Raw and Turbo. The two share + one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" + - so every LoRA tensor shape matches on both and ComfyUI/diffuser + +─────────────────────────────────────────────────────────────────────────────── + PROS +─────────────────────────────────────────────────────────────────────────────── + + Widely adopted — 28,534 downloads. [source: CivitAI stats] + + Positively received — 2,683 thumbs-up, 1 down. [source: CivitAI stats] + + Explicit trigger word(s) — azure youth anime art style — so the effect + can be turned on and off from the prompt. [source: model metadata] + + Author published concrete recommended settings. [source: description] + +─────────────────────────────────────────────────────────────────────────────── + CONS & CAVEATS +─────────────────────────────────────────────────────────────────────────────── + - Author says the LoRA may be overtrained — start at low strength. + [source: description] + - Needs a reduced strength; the default 1.0 is likely too strong. + [source: description] + - Author references NSFW output — review before using in shared work. + [source: description] + - Author never states which generation mode it was trained for; test + against your own workflow before relying on it. [source: description] + - Large at 895 MB — slower to load and heavier to stack. [source: file + size] + +─────────────────────────────────────────────────────────────────────────────── + COMPATIBILITY WARNING +─────────────────────────────────────────────────────────────────────────────── + Silent degradation, never a hard error, between Raw and Turbo. The two + share one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" - so + every LoRA tensor shape matches on both and ComfyUI/diffusers will load an + adapter against either without complaint. The lvladikov 4-step LoRA card + states this for the reverse direction: "Every layer it targets also exists + in Krea 2 Raw, so it will load there without complaint - but that is a side + effect of the shared architecture, not a supported mode", and reports + mixed, subject-dependent results with malformed subjects (duplicated heads, + fused limbs) at low step counts, needing 14+ steps with Raw's CFG on before + output is coherent. The intended direction (train on Raw, run on Turbo) is + officially blessed. The more common practical failure is a settings + mismatch rather than a weight mismatch: running a Turbo-targeted LoRA at + Raw's 52 steps + CFG 3.5, or leaving CFG enabled on Turbo, degrades output + with no warning. Loading a Krea 2 LoRA against a different model family + entirely (Flux, Qwen-Image, Z-Image) is also not an error - ComfyUI logs + "lora key not loaded: {key}" per key and generates as if no LoRA were + applied. + +─────────────────────────────────────────────────────────────────────────────── + AUTHOR'S DESCRIPTION +─────────────────────────────────────────────────────────────────────────────── + KREA 2 V2 Update: Dataset overhaul, trained 10 epoch each on 512/768/1024. + Dim set to 64 for test. + The halo should be easy to remove by reduce LoRA weight. The best balance + is 0.8 in my test. + Trigger word has changed like official Krea 2 loras: + azure youth anime art style But you should be able to use it without + trigger word, cuz it's pretty weak. + It can be used as a high-contrast or flat color style module. + Use proper prompt like Krea 2 prompting guide . Still, the model do not + contain any character information. You still need character lora if you + want a typical character. + ==================== + KREA 2 V1 Update: I've back for a little test, trained on my BAXL dataset + with NL captions. Trained on my own 4090. + Weight 1 is fine, Trigger word is + BArtstyle style USE WITH NATURAL LANGUAGE. DANBOORU TAG IS NOT WORKING. + If it doesn't work well, try remove words like "anime, cartoon" in the + prompt. It may slightly affect the style. + Halo may undesirably come out in certain prompts. If happens, try a + different prompt or reduce lora weight, or just try another seed will help. + This model do not contain any character information. You still need + character lora to make it work properly, this is limited by Krea 2's base + model. + I will test it after I train my first character lora for Krea2 and update + info here. + + Info below has expired. Time flies huh? It has been 2 years since I train + model. + =============================================== + + I've wasted too much time on this model... Probably this will be the final + update, I may focus on SDXL when 4090 be affordable to me, or just fixing + other LoRA I made if 4090 is still cost way too much in next half year. + The original decision was using dreambooth for update, but I'm using 3060 + so the time effort is not good. So, still LoRA. + REv3 Change Log： - Base model now have 3 versions, which have different + feature: + - CF3 - Have a good balance between replica art style and user-friendly. + - You should use it on CounterFeitV3 (and those CF3's mixed model, e.g. + PrimeMix/MixPro) + + - DO NOT USE IT WITH Anything SERIES. I tried and it looks awful. + + - Silicon29 - Trained on Silicon29-Dark, it sacrificed a bit BA's clean + line art, but it has a more BA-ish vivid color+god ray than CF3 version. + - I strongly recommend this version (at least you should try it out!) + + - It has a better variety with non-base model when still kept the base + model's characteristic. + + - Pink theme is its weakness. If you want to generate the BA Finale's blue- + pink gradient scenery, use CF3/ANY version instead. + + - + + - ANY - Trained on ANYv5, harder to use than CF3 or Silicon29, but it seems + to be the most BA one. Only suggest when you using AnythingV5 or those + model based on it. + + - Dataset rehauled + - Manually downscale + denoise before send it to bucket to avoid aliasing. + + - Trained resolution is 1024x768 . So it's better generate horizonal image + than vertical one. Suggest generate resolution : + - 432x768 (Vertical, x2.5 for 1080P) + + - 960x544 (Horizonal, x2 for 1080P) + + - Concept change + - now the typical BA style (blue sky+ modern city / ruined classroom + + water) has its own concept. + - trigger word: blue sky + + - blue theme now can generate nighttime. + + - Preivew (test on Silicon29 ver.) + - + + - Block Weight Training + - Use a more conservative set. So these LoRAs will affect more anatomy but + not sacrificed too much effect. + + - While using REv3, your LoRA weight can up to 1, s ince the anatomy is + basically hopeless XD + - joke aside, v9b has sacrificed too much color for meh anatomy while STILL + CAN NOT use at a higher weight. + + - REv3's anatomy is better than v7 I guess. + + - I've tested some sampler. + - For conclusion: Restart ≈ DPM++ 2M Karras > DPM++ 2S a Karras > any + sampler with SDE + - UniPC is basically a downgrade version of Restart. + + - DPM++ 2S a Karras is oversaturated on typical model and weight. not bad + but just not BA style, you should use a lower weight on this sampler. + + - Restart & DPM++ 2M Karras will cause blue or green spots or something + like jpeg artifact in typical situation. It can not be fixed cuz it's base + model's problem - Although CounterFeit V3 provide a relatively good art, + when still focus on prompts, the model itself has been overtrained (I + believe that's why it's very picky on VAEs.) + + - ↑ that's why you should use CFv2.5's VAE while using this lora models. + + ================== + V9b Change Log： - Base model changed to CounterFeit V3. + + - Use Block Weight Training. Anatomy should be way better than V7. + + - BAPV no longer used as trigger word. You can still use color scheme to + trigger typical color. + + - For the best result, you should choose DPM++ 2S a Karras sampler & use + 4x-ultrasharp as upscaler. + + Known Issue： - Color shifting has been noticeably sacrificed for better + anatomy. Check the last preivew for more info. V9b will affect less on the + whole image's composition and charactor posing. I might release a version + without Block Weight Training, which has better colors, but you should know + how to properly use LoRA Block Weight. + + - It tends to give a high-contrast two-tone shadow even on nighttime. + Consider combine with night concept if you need to generate dark scene. + + - Bad Composition still occurs if you are using a relatively high weight + (over 0.7), consider use LoRA Block Weight (LBW) with OUTALL setting to + reduce the affect. If you use LBW, higher weight up to 1 is Okay. + + V7: Since the training data is based on Blue Archive's PV & CG screenshot, + this LoRA do not very capable with NSFW content. + V9 might be better. But I haven't tried it out. Feel free to test it by + yourself! + Weight 0.6-0.7 is good enough. About 0.4-0.8 is working zone. Higher will + have a better color scheme but worse anatomy. + ================== + Support color scheme in grid: (V7 sample) + + Still, it only works perfectly without any charactor LoRA. You can get + similar result with charactor LoRA, but you need to be patient. V9b has + been MUCH better, but still you need to roll dices. + + The method below is only for older LoHA version. =============== + Compared to Bluemix (the 4g ckpt model), this LoHA is more capable with + character LoRA, it will tend to be keeping character's facial + characteristic and outfit, only change the line and coloring. (like a color + filter rather than totally different art style). + If you like the character be fully switch to BA art style, generating OC or + don't care about the LoRA character's outfit, I suggest you to use BlueMix + instead. + + Suggestion: + - use nature language to generate base environment (e.g. A girl is standing + alongside the river. ) + + - weight 0.8-1, (higher value for better result but worse anatomy, vice + versa.) + + - base resolution: w or h lower than 768x768, over will cause the image + broken more often. + + - base model: suggest anyv3/v5, not recommend any oil painting/pastel model + like AOM series or counterfeit. It CAN use, but is only a color filter. + + - environment prompts (still work!): + - Blue Archive-ish theme (Millennium?): tree, cloud, road, city, day, blue + sky, outdoors + + - Japanese theme (Hyakkiyako?): cloud, day, outdoors, cherry blossoms, + + - Nightsky: night, night sky, starry sky, + + - You can try other tags generated by taggers of Blue Archive CGs. + + === + I'm learning how to make a LoHA/artstyle model. This is my first step + result, my PC isn't very good for testing so I release it for more + feedbacks from your side, so this is basically an Alpha test, the model do + work but not very user-friendly. + If you want to get a good result, you should know how to use lora block + weight. + If you don't know how to use it but still have interest on learning some + new tricks, then read the following guide. My goal is make a artstyle LoHA + that is capable to most of chara LoRAs. + So I trained it on Animefull-pruned, works perfectly on that one and + Anything V3, will work on AnythingV4.5 without character LoRA, but almost + won't work on Counterfeit or AOM3. It will more likely a color filter than + an art style on those models. + Maybe I should train another version based on AnythingV4.5. If you have any + suggestion about art style training, let me know. I will be very appreciate + of that. + + It can work without tags, but sometimes will make weird color. + Here are some tag groups for more stable results: + - Blue Archive-ish theme (Millennium?): tree, cloud, road, city, day, blue + sky, outdoors + + - Japanese theme (Hyakkiyako?): pink theme, cloud, day, outdoors, cherry + blossoms, + + - Nightsky: blue theme, night, night sky, starry sky, cityscape, + + - You can try other tags generated by taggers of Blue Archive CGs. + + You 'd better use it with LoRA block weight if you want to input a + character lora. You can get it as webui plugin. Otherwise you should lower + this model's weight to below 0.6, which will work bit poorly but still + work, or it will make the image color very awful. + ↓If you get something like this, lower the artstyle model weight, or use + the method below. + + When you use it with LoRA block weight, there are 2 methods: + - A: By default, you should remove your charactor lora's artstyle by this: + + don't forget to add :ALL to this model or it may won't be working. + e.g. + + - B: If you noticed that your character's appearance has been obviously + affected by this model, keep only the color of the artstyle model instead + like this: + + In this situation, you can keep Charactor model in full. + e.g. + + - Basically, method A will works more like the original artstyle than + method B. + + - It seems that LoHA model have more block weights, I don't have much time + to test since my PC will easily crash on it, if you have some information + about it, please tell me. But so far using LoRA block weight is fine. + + - ↓Still, if you get everything right but still got something like this, + try to use the second way instead. + + Method A when the charactor is affected. + + Method B + +─────────────────────────────────────────────────────────────────────────────── + VERSION NOTES +─────────────────────────────────────────────────────────────────────────────── + Dataset overhaul + +─────────────────────────────────────────────────────────────────────────────── + EXAMPLE MEDIA +─────────────────────────────────────────────────────────────────────────────── + 6 preview(s) in ./example_images/ — community results for + this LoRA. Videos keep a matching _poster.jpg still frame. + + No prompts are recorded below: CivitAI's public API flags these posts as + having generation metadata but does not return it, so the prompts cannot be + scraped. Open the model page to read them. + + 01_image.jpeg (image, 1288x1936) + 02_image.jpeg (image, 1536x2728) + 03_image.jpeg (image, 1536x2728) + 04_image.jpeg (image, 1024x1024) + 05_image.jpeg (image, 1536x2728) + 06_image.jpeg (image, 1288x1936) + +─────────────────────────────────────────────────────────────────────────────── + HOW TO USE (COMFYUI) +─────────────────────────────────────────────────────────────────────────────── + 1. Update ComfyUI to at least v0.26.0 (released 2026-06-23), which added + local Krea2 support. For reference-image / style-reference / edit LoRAs you + need at least v0.29.0 (2026-07-29), which added regular and timestep-zero + reference images for Krea 2. ModelMergeKrea2 arrived in v0.27.0 + (2026-06-30). + + 2. Put the base stack in place (files from huggingface.co/Comfy-Org/Krea-2, + which is not gated): + - ComfyUI/models/diffusion_models/krea2_turbo_fp8_scaled.safetensors + (documented recommended default; the repo also holds krea2_turbo_bf16, + _int8_convrot, _mxfp8, _nvfp4 and krea2_raw_bf16 / _fp8_scaled / + _int8_convrot) + - ComfyUI/models/text_encoders/qwen3vl_4b_fp8_scaled.safetensors (bf16 + also available) + - ComfyUI/models/vae/qwen_image_vae.safetensors + + 3. Put the LoRA .safetensors in ComfyUI/models/loras/ - flat, no per-model + subfolder required. + + 4. Build or load the graph. The official image_krea2_turbo_t2i.json + template is: UNETLoader (krea2_turbo_fp8_scaled.safetensors) -> + LoraLoaderModelOnly -> KSampler; CLIPLoader + (qwen3vl_4b_fp8_scaled.safetensors, type "krea2") -> text encode -> + KSampler positive, with ConditioningZeroOut into negative; VAELoader + (qwen_image_vae.safetensors) -> VAEDecode -> SaveImage; EmptyLatentImage + 1024x1024 driven by ResolutionSelector. An int8 sibling template + (image_krea2_turbo_t2i_int8.json) also ships. + + 5. Load the LoRA with LoraLoaderModelOnly, not LoraLoader - both official + templates do. Krea 2 LoRAs patch the DiT only; note that the DiT's own + text_fusion / txtfusion blocks are a normal LoRA target (the official style + LoRAs patch them), but the Qwen3-VL text encoder itself is not. Set + strength 1.0 for Krea's official style LoRAs (the shipped template uses + 0.8) and add the LoRA's trigger word to the prompt. + + 6. Sampler settings: steps 8, cfg 1, sampler euler, scheduler simple, + denoise 1.0. Do not add a ModelSampling node for plain t2i - ComfyUI + already applies shift 1.15 from Krea2.sampling_settings. + + 7. For style reference, use the dedicated template: UNETLoader + krea2_turbo_int8_convrot.safetensors, LoraLoaderModelOnly + krea2_style_reference.safetensors at 1.0, ModelSamplingFlux (1.15, 0.5), + TextEncodeQwenImageEditPlus fed by LoadImage, + FluxKontextMultiReferenceLatentMethod set to index_timestep_zero, and + SamplerCustomAdvanced with KSamplerSelect euler + BasicScheduler simple/8 + + CFGGuider cfg 1. + + 8. For instruction editing (Krea 2 Identity Edit), stock nodes are not + enough: install the ComfyUI-Krea2Edit node pack, which ships its own + workflows, and run Turbo at 8-12 steps / cfg 1 (or Raw at 20 steps / cfg 3 + for removals) with LoRA strength 1.0. + + 9. Low-VRAM alternative: install city96's ComfyUI-GGUF and use a community + GGUF build (realrebelai/KREA-2_GGUFs, quantized from Turbo; + molbal/krea2-gguf, quantized from Raw) with the Unet Loader (GGUF) node in + place of UNETLoader. There is no official GGUF release. LoRA loading is + unchanged. + +─────────────────────────────────────────────────────────────────────────────── + Scraped by lorakit on 2026-09-05 00:11 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/bartstyle-blue-archive-art-style-lora-anime-flat-color/metadata.json b/bartstyle-blue-archive-art-style-lora-anime-flat-color/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..728442164e0c069d13db50e3eb4af947c883ae91 --- /dev/null +++ b/bartstyle-blue-archive-art-style-lora-anime-flat-color/metadata.json @@ -0,0 +1,144 @@ +{ + "name": "BArtstyle | Blue Archive Art Style LoRA | Anime Flat Color | 蔚蓝档案画风模型 | ブルーアーカイブ画風モデル", + "slug": "bartstyle-blue-archive-art-style-lora-anime-flat-color", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 29215, + "version_id": 3220090, + "version_name": "Krea2-V2", + "source": "civitai", + "url": "https://civitai.com/models/29215", + "repo_id": null, + "file_path": null, + "author": "4BEraser", + "category": { + "key": "style", + "label": "Visual style" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [ + "azure youth anime art style" + ], + "recommended_settings": { + "LoRA strength": "0.6–0.7", + "Resolution": "1024x768", + "Sampler": "DPM++ 2M" + }, + "chaining": { + "author": { + "stacks": "not stated", + "notes": "", + "avoid": [], + "requires": [], + "order": "", + "strength": "", + "stated": false + }, + "general_rules": [ + "The most important rule is not about chaining but about which checkpoint you chain onto: the official README says in bold \"TRAIN on Raw and RUN on Turbo\", and \"We highly recommend using RAW for training LoRAs and applying them on Turbo for inference.\" LoRAs trained on Raw are stated to work well on Turbo.", + "Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 LoRAs are DiT-only, so there is no CLIP side to chain and the text-encoder branch stays untouched however many LoRAs you add.", + "Krea 2 LoRAs stack in practice and the community treats multi-LoRA as routine - style + character, style + slider, realism + character. Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one writes \"I recently stacked over 20 of my own LoRAs on a single generation and the output still came out great.\"", + "Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is loaded. The CivitAI 12GB training guide also notes rank affects composability: \"rank 8 produced a lighter style influence that was easier to stack with other LoRAs, rank 16 provided a useful balance between concept strength and composability, and rank 32 was better suited to a precise object or concept intended to take priority over other stacked LoRAs.\"", + "Accelerator LoRAs are position-sensitive in intent, not in node order: a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to-Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is explicitly recommended by the 4-step LoRA's author.", + "Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card says \"Keep mu = 1.15. The training targets are anchored to that grid; a different shift moves the [sampler off them]\"; the same reasoning applies to any Turbo-trained adapter in the chain.", + "Reference/edit LoRAs are designed to be stacked under ordinary subject and style LoRAs - the Identity Edit card states \"Composes with your LoRAs: character/body/style LoRAs stack on top and steer the prior\" and recommends a stacked subject LoRA when unusual subjects drift toward the base prior.", + "Krea 2 LoRAs cannot be chained with LoRAs for any other base model. Many CivitAI pages carry multi-model titles like \"(Krea-2 + ZIT)\" or \"[Flux | ZIB | Krea2]\" - those are separate files for separate architectures shipped under one model page, not a combined adapter.", + "To combine two full Krea 2 checkpoints rather than adapters, ComfyUI ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per-section weights for first., tmlp., txtmlp., tproj., txtfusion.projector, txtfusion.layerwise_blocks.0-1, txtfusion.refiner_blocks.0-1, blocks.0-27 and last.", + "The author never states which weight family this targets. Silent degradation, never a hard error, between Raw and Turbo. The two share one architecture - ai-toolkit's config is commented \"The reference 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)\" - so every LoRA tensor shape matches on both and ComfyUI/diffuser" + ] + }, + "pros": [ + "Widely adopted — 28,534 downloads. [source: CivitAI stats]", + "Positively received — 2,683 thumbs-up, 1 down. [source: CivitAI stats]", + "Explicit trigger word(s) — azure youth anime art style — so the effect can be turned on and off from the prompt. [source: model metadata]", + "Author published concrete recommended settings. [source: description]" + ], + "cons": [ + "Author says the LoRA may be overtrained — start at low strength. [source: description]", + "Needs a reduced strength; the default 1.0 is likely too strong. [source: description]", + "Author references NSFW output — review before using in shared work. [source: description]", + "Author never states which generation mode it was trained for; test against your own workflow before relying on it. [source: description]", + "Large at 895 MB — slower to load and heavier to stack. [source: file size]" + ], + "file": { + "name": "BArtstyle-KREA2-V2-1024.safetensors", + "size_bytes": 938539208, + "size_mb": 895.06, + "sha256": "9c19dc382d92002bfc53ece7df5556acc288579af9f0719752059dfae7fe9f70" + }, + "stats": { + "downloads": 28534, + "thumbs_up": 2683, + "thumbs_down": 1 + }, + "popularity": { + "downloads": 28534, + "thumbs_up": 2683, + "thumbs_down": 1, + "comments": 0, + "approval": "100% positive (2683/2684)", + "age_days": 24.2, + "downloads_per_day": 82.34, + "rank_by_downloads": 21, + "rank_by_momentum": 34, + "of_total": 150, + "adoption_tier": "High", + "momentum": "Trending · fresh release" + }, + "selection_score": 61130.0, + "selection_reasons": [ + "28,534 downloads", + "2,683 thumbs-up", + "documented triggers + usage notes", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "1288x1936", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "1536x2728", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "1536x2728", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "1024x1024", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "1536x2728", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "1288x1936", + "prompt": "" + } + ], + "tags": [ + "style", + "game art", + "artstyle", + "blue archive", + "anime" + ], + "scraped_at": "2026-09-05T00:11:48+00:00", + "path": "krea-2/bartstyle-blue-archive-art-style-lora-anime-flat-color" +} diff --git a/famegrid-krea-2-z-image-qwen-realism-lora/example_images/01_image.jpeg b/famegrid-krea-2-z-image-qwen-realism-lora/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..0c0944b6c94d3c2384cca216fde2bf43f4f9d56f --- /dev/null +++ b/famegrid-krea-2-z-image-qwen-realism-lora/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:58e6028a4fceee2de846fcf6e4ae0758e99d695ae8cf68a8a59cebbf492fa833 +size 681738 diff --git a/famegrid-krea-2-z-image-qwen-realism-lora/example_images/02_image.jpeg b/famegrid-krea-2-z-image-qwen-realism-lora/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..12f6f29c9b6552249f2ef49b23815fbfb23d0b92 --- /dev/null +++ b/famegrid-krea-2-z-image-qwen-realism-lora/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:67651a37c9113e59ff46b5a25a3563e927ef090ebb401f845749ebf0a9478b5b +size 772812 diff --git a/famegrid-krea-2-z-image-qwen-realism-lora/example_images/03_image.jpeg b/famegrid-krea-2-z-image-qwen-realism-lora/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..3bbc094dd77727244811561dae9480d14eff8fdb --- /dev/null +++ b/famegrid-krea-2-z-image-qwen-realism-lora/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2463c661b4fd113faff28d3ebff2be4bf654c5845def9b3c2cb735e953ab9c0d +size 4057506 diff --git a/famegrid-krea-2-z-image-qwen-realism-lora/example_images/04_image.jpeg b/famegrid-krea-2-z-image-qwen-realism-lora/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..0b43acce89b1abc900f5a3964810fd6e8a7c06b8 --- /dev/null +++ b/famegrid-krea-2-z-image-qwen-realism-lora/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dfc0b36d6c48fc6c689c45a9cd0b9aa4abfe3a291398ee76f86ea794e6abc036 +size 4009396 diff --git a/famegrid-krea-2-z-image-qwen-realism-lora/example_images/05_image.jpeg b/famegrid-krea-2-z-image-qwen-realism-lora/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..eb6b7c8f640fa9c73242926ef95d6794ca8eb7b3 --- /dev/null +++ b/famegrid-krea-2-z-image-qwen-realism-lora/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:688e0e0b5ec7a53f5fd93837b53bcdfd1b3721a304e001ea02c549225741257d +size 934764 diff --git a/famegrid-krea-2-z-image-qwen-realism-lora/example_images/06_image.jpeg b/famegrid-krea-2-z-image-qwen-realism-lora/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..487adcc9b79c41a967c18d63d9d49ee92d22e971 --- /dev/null +++ b/famegrid-krea-2-z-image-qwen-realism-lora/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b16061cff4398ca9b4ee05e3848057c4a7fb88dfb0f6a2379b2de4a5634ab432 +size 738602 diff --git a/famegrid-krea-2-z-image-qwen-realism-lora/famegrid_spicy.safetensors b/famegrid-krea-2-z-image-qwen-realism-lora/famegrid_spicy.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..fbe6d960da39815c94fc38fa47e7be6c8bcdd4f5 --- /dev/null +++ b/famegrid-krea-2-z-image-qwen-realism-lora/famegrid_spicy.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7b4a37a9b9ec3885bf39242f2d4f691f186d959cbaaa30b0630e4631767b425e +size 1562410408 diff --git a/famegrid-krea-2-z-image-qwen-realism-lora/info.txt b/famegrid-krea-2-z-image-qwen-realism-lora/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..b8bf1d5270491692963d3cadedd18104cd038e11 --- /dev/null +++ b/famegrid-krea-2-z-image-qwen-realism-lora/info.txt @@ -0,0 +1,396 @@ +═══════════════════════════════════════════════════════════════════════════════ + FAMEGRID (KREA 2 / Z-IMAGE / QWEN) - REALISM LORA +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode Instruction / reference editing (community, not official) + Checkpoint Krea 2 RAW (krea/Krea-2-Raw) + Category Photorealism + Author UltraMuse + Source https://civitai.com/models/2088956 + Provenance CivitAI · model 2088956 · version 3278885 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File famegrid_spicy.safetensors + Version Krea 2 Spice Preview FIX + Size 1,490.0 MB + Format SafeTensor / unknown precision + SHA256 7b4a37a9b9ec3885bf39242f2d4f691f186d959cbaaa30b0630e4631767b425e + Published 2026-08-29 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + › "famegrid" + +─────────────────────────────────────────────────────────────────────────────── + RECOMMENDED SETTINGS +─────────────────────────────────────────────────────────────────────────────── + Taken verbatim from the author's description: + + LoRA strength 0.3–0.6 + Sampler steps 8–16 + CFG / guidance 1.0 + +─────────────────────────────────────────────────────────────────────────────── + ADOPTION & POPULARITY +─────────────────────────────────────────────────────────────────────────────── + Downloads 39,677 + Rating 100% positive (1818/1826) + Comments 89 + Published 6 days ago + Download rate 522.5/day since release + + Adoption rank #11 of 150 in this collection (Top tier) + Momentum rank #2 of 150 by download rate (Hot · fresh release) + + Adoption tier is the percentile of total downloads within this collection; + momentum is the percentile of downloads-per-day since release. Momentum is + ranked rather than measured against a fixed rate, because this base model + is itself new and every LoRA looks fast on an absolute scale. Both numbers + are CivitAI's, read on the scrape date at the foot of this file. + +─────────────────────────────────────────────────────────────────────────────── + CHAINING / STACKING +─────────────────────────────────────────────────────────────────────────────── + Stacks with others not stated + + The author gave no chaining guidance. The rules below are the + general Krea 2 ones, not anything specific to this LoRA. + + General Krea 2 rules that apply here (see REFERENCE.txt): + + • The most important rule is not about chaining but about which + checkpoint you chain onto: the official README says in bold "TRAIN on + Raw and RUN on Turbo", and "We highly recommend using RAW for training + LoRAs and applying them on Turbo for inference." LoRAs trained on Raw + are stated to work well on Turbo. + • Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 + LoRAs are DiT-only, so there is no CLIP side to chain and the text- + encoder branch stays untouched however many LoRAs you add. + • Krea 2 LoRAs stack in practice and the community treats multi-LoRA as + routine - style + character, style + slider, realism + character. + Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one + writes "I recently stacked over 20 of my own LoRAs on a single + generation and the output still came out great." + • Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge + on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is + loaded. The CivitAI 12GB training guide also notes rank affects + composability: "rank 8 produced a lighter style influence that was + easier to stack with other LoRAs, rank 16 provided a useful balance + between concept strength and composability, and rank 32 was better + suited to a precise object or concept intended to take priority over + other stacked LoRAs." + • Accelerator LoRAs are position-sensitive in intent, not in node order: + a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a + sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to- + Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is + explicitly recommended by the 4-step LoRA's author. + • Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI + (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card + says "Keep mu = 1.15. The training targets are anchored to that grid; a + different shift moves the [sampler off them]"; the same reasoning + applies to any Turbo-trained adapter in the chain. + • Reference/edit LoRAs are designed to be stacked under ordinary subject + and style LoRAs - the Identity Edit card states "Composes with your + LoRAs: character/body/style LoRAs stack on top and steer the prior" and + recommends a stacked subject LoRA when unusual subjects drift toward + the base prior. + • Krea 2 LoRAs cannot be chained with LoRAs for any other base model. + Many CivitAI pages carry multi-model titles like "(Krea-2 + ZIT)" or + "[Flux | ZIB | Krea2]" - those are separate files for separate + architectures shipped under one model page, not a combined adapter. + • To combine two full Krea 2 checkpoints rather than adapters, ComfyUI + ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per- + section weights for first., tmlp., txtmlp., tproj., + txtfusion.projector, txtfusion.layerwise_blocks.0-1, + txtfusion.refiner_blocks.0-1, blocks.0-27 and last. + • Undistilled base checkpoint, no step or CFG distillation and no post- + training. The card says "its not recommended for inference use. Its a + good base for finetuning or post-training". Official settings: 52 + steps, CFG 3.5, trained up to 1K. This is where Krea tells you to train + LoRAs. + +─────────────────────────────────────────────────────────────────────────────── + PROS +─────────────────────────────────────────────────────────────────────────────── + + Widely adopted — 39,677 downloads. [source: CivitAI stats] + + Positively received — 1,818 thumbs-up, 8 down. [source: CivitAI stats] + + Explicit trigger word(s) — famegrid — so the effect can be turned on + and off from the prompt. [source: model metadata] + + Author published concrete recommended settings. [source: description] + +─────────────────────────────────────────────────────────────────────────────── + CONS & CAVEATS +─────────────────────────────────────────────────────────────────────────────── + - Author labels this experimental — results vary shot to shot. [source: + description] + - Needs a reduced strength; the default 1.0 is likely too strong. + [source: description] + - Author references NSFW output — review before using in shared work. + [source: description] + - Large at 1490 MB — slower to load and heavier to stack. [source: file + size] + +─────────────────────────────────────────────────────────────────────────────── + COMPATIBILITY WARNING +─────────────────────────────────────────────────────────────────────────────── + Silent degradation, never a hard error, between Raw and Turbo. The two + share one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" - so + every LoRA tensor shape matches on both and ComfyUI/diffusers will load an + adapter against either without complaint. The lvladikov 4-step LoRA card + states this for the reverse direction: "Every layer it targets also exists + in Krea 2 Raw, so it will load there without complaint - but that is a side + effect of the shared architecture, not a supported mode", and reports + mixed, subject-dependent results with malformed subjects (duplicated heads, + fused limbs) at low step counts, needing 14+ steps with Raw's CFG on before + output is coherent. The intended direction (train on Raw, run on Turbo) is + officially blessed. The more common practical failure is a settings + mismatch rather than a weight mismatch: running a Turbo-targeted LoRA at + Raw's 52 steps + CFG 3.5, or leaving CFG enabled on Turbo, degrades output + with no warning. Loading a Krea 2 LoRA against a different model family + entirely (Flux, Qwen-Image, Z-Image) is also not an error - ComfyUI logs + "lora key not loaded: {key}" per key and generates as if no LoRA were + applied. + +─────────────────────────────────────────────────────────────────────────────── + AUTHOR'S DESCRIPTION +─────────────────────────────────────────────────────────────────────────────── + 🚀 New Release — Krea 2 Spice Preview Preview Fix update: Apologies for the + screw-up; I uploaded the wrong checkpoint in the last release. Replace that + one with this here. WORKFLOW: https://pastebin.com/GKMm2Njs Still use the + original model page to upload to the gallery ---------- + This preview model was trained on approximately 1/20th of the dataset + planned for the final release. It builds on FameGrid Natural’s realistic, + everyday photographic style, producing casual images that look like + authentic modern smartphone photos rather than polished or heavily staged + shoots. Important: If using with character lora keep strength 0.3-0.6. You + can also try turning down character lora as much as possible to keep + likeness and turn up the Famegrid lora more. + The preview currently has minimal NSFW support involving men and works with + the same workflow as FameGrid Natural V1. + Contribute to training of the full version. Cloud GPU training has become + too expensive for me to continue funding these models alone. The final + version will therefore be released as an exclusive model for contributors + only (as crowdfunding didn't work). + Contributions over $10 will receive the final model once training is + complete. After the funding threshold is reached, local access will be + available exclusively to qualifying contributors—there will be no other way + to obtain the model for local use. + The model will also be hosted on FameGrid.ai. + Secure your contributor access here: + https://ultramuse.art/train-lora/fundraiser + Very limited contributor spaces are available. + 🔑 Trigger Word Famegrid Start your prompt with the trigger word. + Example: + Famegrid, casual smartphone photo of a woman sitting at a neighborhood + coffee shop ⚙️ Recommended LoRA Weight Recommended range: 0.3–1.0 + - Use a lower weight for a subtle influence. + + - Use a higher weight for a stronger FameGrid Natural aesthetic. + + - Adjust within the recommended range to suit your prompt and workflow. + + 📥 Workflow Available on Discord The FameGrid Natural Krea 2 workflow is + available for free in the Free Resources channel of the UltraMuse Discord. + 👉 Join the UltraMuse Discord and Download the Workflow 👉 You can also use + this model and train LoRAs at Famegrid.ai + + 🎬 Introducing Creator Studio The All-in-One AI Influencer Operating System + Creator Studio brings the entire AI influencer workflow into one connected + production studio. + Plan content, create consistent characters, generate and edit images, + produce videos and audio, write captions, publish posts, and review + performance—all from one creator-focused platform. + We tested more than 100 AI models and platforms, selected the strongest + options, and put them inside one useful wrapper with the tools creators + need to produce high-quality AI content. + Whether you manage one virtual creator or an entire roster, Creator Studio + helps turn scattered generations into a repeatable content-production + system. + Bring Your Own API or ComfyUI Workflow Creator Studio is not a closed + ecosystem. Add any API or import a custom ComfyUI workflow, give the agent + the documentation, and it can figure out the integration, build the + controls, and generate the UI for it. + All 27 Creator Studio Tools Create and Direct • Copilot • Image Generation + • Image Edit • AI Swap • Copy + Characters and Campaigns • Photoshoot Studio • Character Creation • + Character Manager + Video and Sound • Video Studio • Motion Control • Audio Studio • Lipsync + Studio + Finish and Enhance • Upscale Studio • Skin Enhancement • Post-Process + Workflows and Operations • ComfyUI Workflows • Provider Builder • Caption + Studio • Publish Studio • Social Analytics • LoRA Training + Automated Flows • Inspiration Dump • InstaClone • InstaReel • UGC Creator • + Hybrid Lipsync • Watermark + See It in Action https://ultramuse.art/creatorstudio/demo/ + ━━━━━━━━━━━━━━━━━━━━ 🧬 Krea 2 & Z-Image LoRA Trainer Train your own custom + character and style LoRAs for Krea 2 and Z-Image . + The trainer helps you prepare your dataset, configure the training, monitor + its progress, and create reusable LoRAs for consistent AI influencers, + characters, styles, and creator content. + The Krea 2 and Z-Image LoRA Trainer is available through the UltraMuse + Discord server: + 👉 [Access the LoRA Trainer Here](https://ultramuse.art/train-lora] + 🚀 Newest Release — July 2026 🔥 FameGrid Krea 2 Standard v1 The Highest- + Quality FameGrid Yet FameGrid Krea 2 is the highest-quality FameGrid + release yet. It produces excellent social-media-style images with natural + posing, believable skin, polished lighting, strong compositions, and the + modern creator aesthetic FameGrid was built for. + This release is designed for AI influencers, lifestyle content, fashion and + beauty photography, creator campaigns, product promotions, and premium + social posts. It delivers cleaner, more convincing results while keeping + images visually engaging and ready for a feed. + What It Does • Creates premium social-media and creator-style imagery + • Produces natural poses and believable body language + • Improves skin, lighting, styling, and overall visual polish + • Excels at lifestyle, fashion, beauty, influencer, and promotional content + • Maintains a clean, modern aesthetic without overly complicated prompts + Trigger Word famegrid + Start your prompt with the trigger word. Example: + famegrid, candid social media photo of... + Recommended Starting Settings • Sampler: er_sde + • Guidance: 1 + • Steps: 8–16 + These are starting points. Adjust the LoRA weight, steps, and workflow to + suit the look you want. + Best Uses ✨ Social media posts • AI influencers and virtual creators • + Lifestyle photography • Fashion and beauty content • Product promotions • + Marketing campaigns • Character-driven content + Workflow Available on Discord The official FameGrid Krea 2 workflow is + available through the UltraMuse Discord, along with updates, tutorials, + support, and additional resources. + https://discord.gg/4zbGm5j6jW 🚀 Support FameGrid Krea 2 V2 (Enhanced NSFW) + Help fund the next version of FameGrid Krea 2 , including improved realism, + anatomy, posing, prompt adherence, social-media image quality, and expanded + NSFW support. + 👉 [Support FameGrid Krea 2 V2]( https://ultramuse.art/train- + lora/fundraiser) + ━━━━━━━━━━━━━━━━━━━━ + 📚 Previous FameGrid Releases 🔥 FameGrid Spice V2 — June 2026 Spice V2 + delivered a major jump in skin quality, posing, prompt following, body-type + variety, male generation quality, ethnicity representation, and artifact + reduction. + Recommended Spice V2 Settings • Sampler: sa_diag_implicit / qin_zhang_2s + • Scheduler: beta57 + • Steps: 9 (20 with the ClownShark BetaSamplingScheduler) + • CFG: 1.0 • ETA: 0.50 • Denoise: 1.0 + • Sampler Mode: resample • Bongmath: true + 🚀 FameGrid Revolution The previous Z-Image flagship, compatible with + Z-Image Base and Turbo and built for strong prompt adherence, influencer + photography, and social-media aesthetics. + Available variants include FameGrid Revolution Bold, More Bold, ExtraReal, + and UltraReal. + Model and workflow: + https://civitai.com/models/2431502/famegrid-revolution-hybrid-workflow-zib- + zit + 🎨 FameGrid Qwen Series The FameGrid Qwen releases introduced improved skin + realism and more believable subjects. + Trigger Words IGMODEL — influencer-style realism + rlskn — enhanced skin texture and natural imperfections + Recommended Qwen Settings • Model Flow: 2 • Sampler: res_2s • Scheduler: + simple + • CFG: 1–3 (1 recommended) • LoRA Weight: 1.0–2.0 (1.0 recommended) + ━━━━━━━━━━━━━━━━━━━━ + 💬 Community and Resources Get official workflows, updates, checkpoints, + tutorials, prompt resources, troubleshooting help, experimental releases, + community showcases, and early access in the UltraMuse Discord. + Discord: https://discord.gg/4zbGm5j6jW + UltraMuse: https://ultramuse.art + ━━━━━━━━━━━━━━━━━━━━ + ❤️ About FameGrid FameGrid was built around one goal: to create modern + social-media-quality imagery that feels believable, natural, and visually + compelling. + The focus has always been natural posing, realistic skin, clean + compositions, modern lighting, strong prompt adherence, and professional + creator aesthetics. + +─────────────────────────────────────────────────────────────────────────────── + EXAMPLE MEDIA +─────────────────────────────────────────────────────────────────────────────── + 6 preview(s) in ./example_images/ — community results for + this LoRA. Videos keep a matching _poster.jpg still frame. + + No prompts are recorded below: CivitAI's public API flags these posts as + having generation metadata but does not return it, so the prompts cannot be + scraped. Open the model page to read them. + + 01_image.jpeg (image, 1248x1824) + 02_image.jpeg (image, 1248x1824) + 03_image.jpeg (image, 1200x1600) + 04_image.jpeg (image, 1200x1600) + 05_image.jpeg (image, 1248x1824) + 06_image.jpeg (image, 1248x1824) + +─────────────────────────────────────────────────────────────────────────────── + HOW TO USE (COMFYUI) +─────────────────────────────────────────────────────────────────────────────── + 1. Update ComfyUI to at least v0.26.0 (released 2026-06-23), which added + local Krea2 support. For reference-image / style-reference / edit LoRAs you + need at least v0.29.0 (2026-07-29), which added regular and timestep-zero + reference images for Krea 2. ModelMergeKrea2 arrived in v0.27.0 + (2026-06-30). + + 2. Put the base stack in place (files from huggingface.co/Comfy-Org/Krea-2, + which is not gated): + - ComfyUI/models/diffusion_models/krea2_turbo_fp8_scaled.safetensors + (documented recommended default; the repo also holds krea2_turbo_bf16, + _int8_convrot, _mxfp8, _nvfp4 and krea2_raw_bf16 / _fp8_scaled / + _int8_convrot) + - ComfyUI/models/text_encoders/qwen3vl_4b_fp8_scaled.safetensors (bf16 + also available) + - ComfyUI/models/vae/qwen_image_vae.safetensors + + 3. Put the LoRA .safetensors in ComfyUI/models/loras/ - flat, no per-model + subfolder required. + + 4. Build or load the graph. The official image_krea2_turbo_t2i.json + template is: UNETLoader (krea2_turbo_fp8_scaled.safetensors) -> + LoraLoaderModelOnly -> KSampler; CLIPLoader + (qwen3vl_4b_fp8_scaled.safetensors, type "krea2") -> text encode -> + KSampler positive, with ConditioningZeroOut into negative; VAELoader + (qwen_image_vae.safetensors) -> VAEDecode -> SaveImage; EmptyLatentImage + 1024x1024 driven by ResolutionSelector. An int8 sibling template + (image_krea2_turbo_t2i_int8.json) also ships. + + 5. Load the LoRA with LoraLoaderModelOnly, not LoraLoader - both official + templates do. Krea 2 LoRAs patch the DiT only; note that the DiT's own + text_fusion / txtfusion blocks are a normal LoRA target (the official style + LoRAs patch them), but the Qwen3-VL text encoder itself is not. Set + strength 1.0 for Krea's official style LoRAs (the shipped template uses + 0.8) and add the LoRA's trigger word to the prompt. + + 6. Sampler settings: steps 8, cfg 1, sampler euler, scheduler simple, + denoise 1.0. Do not add a ModelSampling node for plain t2i - ComfyUI + already applies shift 1.15 from Krea2.sampling_settings. + + 7. For style reference, use the dedicated template: UNETLoader + krea2_turbo_int8_convrot.safetensors, LoraLoaderModelOnly + krea2_style_reference.safetensors at 1.0, ModelSamplingFlux (1.15, 0.5), + TextEncodeQwenImageEditPlus fed by LoadImage, + FluxKontextMultiReferenceLatentMethod set to index_timestep_zero, and + SamplerCustomAdvanced with KSamplerSelect euler + BasicScheduler simple/8 + + CFGGuider cfg 1. + + 8. For instruction editing (Krea 2 Identity Edit), stock nodes are not + enough: install the ComfyUI-Krea2Edit node pack, which ships its own + workflows, and run Turbo at 8-12 steps / cfg 1 (or Raw at 20 steps / cfg 3 + for removals) with LoRA strength 1.0. + + 9. Low-VRAM alternative: install city96's ComfyUI-GGUF and use a community + GGUF build (realrebelai/KREA-2_GGUFs, quantized from Turbo; + molbal/krea2-gguf, quantized from Raw) with the Unet Loader (GGUF) node in + place of UNETLoader. There is no official GGUF release. LoRA loading is + unchanged. + +─────────────────────────────────────────────────────────────────────────────── + Scraped by lorakit on 2026-09-05 00:11 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/famegrid-krea-2-z-image-qwen-realism-lora/metadata.json b/famegrid-krea-2-z-image-qwen-realism-lora/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..66a85eb2f256faf5756fa496c2e64f959471b654 --- /dev/null +++ b/famegrid-krea-2-z-image-qwen-realism-lora/metadata.json @@ -0,0 +1,152 @@ +{ + "name": "Famegrid (Krea 2 / Z-Image / Qwen) - Realism LoRA", + "slug": "famegrid-krea-2-z-image-qwen-realism-lora", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 2088956, + "version_id": 3278885, + "version_name": "Krea 2 Spice Preview FIX", + "source": "civitai", + "url": "https://civitai.com/models/2088956", + "repo_id": null, + "file_path": null, + "author": "UltraMuse", + "category": { + "key": "photoreal", + "label": "Photorealism" + }, + "modalities": [ + "edit" + ], + "primary_modality": "edit", + "checkpoint_family": "raw", + "trigger_words": [ + "famegrid" + ], + "recommended_settings": { + "LoRA strength": "0.3–0.6", + "Sampler steps": "8–16", + "CFG / guidance": "1.0" + }, + "chaining": { + "author": { + "stacks": "not stated", + "notes": "", + "avoid": [], + "requires": [], + "order": "", + "strength": "", + "stated": false + }, + "general_rules": [ + "The most important rule is not about chaining but about which checkpoint you chain onto: the official README says in bold \"TRAIN on Raw and RUN on Turbo\", and \"We highly recommend using RAW for training LoRAs and applying them on Turbo for inference.\" LoRAs trained on Raw are stated to work well on Turbo.", + "Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 LoRAs are DiT-only, so there is no CLIP side to chain and the text-encoder branch stays untouched however many LoRAs you add.", + "Krea 2 LoRAs stack in practice and the community treats multi-LoRA as routine - style + character, style + slider, realism + character. Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one writes \"I recently stacked over 20 of my own LoRAs on a single generation and the output still came out great.\"", + "Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is loaded. The CivitAI 12GB training guide also notes rank affects composability: \"rank 8 produced a lighter style influence that was easier to stack with other LoRAs, rank 16 provided a useful balance between concept strength and composability, and rank 32 was better suited to a precise object or concept intended to take priority over other stacked LoRAs.\"", + "Accelerator LoRAs are position-sensitive in intent, not in node order: a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to-Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is explicitly recommended by the 4-step LoRA's author.", + "Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card says \"Keep mu = 1.15. The training targets are anchored to that grid; a different shift moves the [sampler off them]\"; the same reasoning applies to any Turbo-trained adapter in the chain.", + "Reference/edit LoRAs are designed to be stacked under ordinary subject and style LoRAs - the Identity Edit card states \"Composes with your LoRAs: character/body/style LoRAs stack on top and steer the prior\" and recommends a stacked subject LoRA when unusual subjects drift toward the base prior.", + "Krea 2 LoRAs cannot be chained with LoRAs for any other base model. Many CivitAI pages carry multi-model titles like \"(Krea-2 + ZIT)\" or \"[Flux | ZIB | Krea2]\" - those are separate files for separate architectures shipped under one model page, not a combined adapter.", + "To combine two full Krea 2 checkpoints rather than adapters, ComfyUI ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per-section weights for first., tmlp., txtmlp., tproj., txtfusion.projector, txtfusion.layerwise_blocks.0-1, txtfusion.refiner_blocks.0-1, blocks.0-27 and last.", + "Undistilled base checkpoint, no step or CFG distillation and no post-training. The card says \"its not recommended for inference use. Its a good base for finetuning or post-training\". Official settings: 52 steps, CFG 3.5, trained up to 1K. This is where Krea tells you to train LoRAs." + ] + }, + "pros": [ + "Widely adopted — 39,677 downloads. [source: CivitAI stats]", + "Positively received — 1,818 thumbs-up, 8 down. [source: CivitAI stats]", + "Explicit trigger word(s) — famegrid — so the effect can be turned on and off from the prompt. [source: model metadata]", + "Author published concrete recommended settings. [source: description]" + ], + "cons": [ + "Author labels this experimental — results vary shot to shot. [source: description]", + "Needs a reduced strength; the default 1.0 is likely too strong. [source: description]", + "Author references NSFW output — review before using in shared work. [source: description]", + "Large at 1490 MB — slower to load and heavier to stack. [source: file size]" + ], + "file": { + "name": "famegrid_spicy.safetensors", + "size_bytes": 1562410408, + "size_mb": 1490.03, + "sha256": "7b4a37a9b9ec3885bf39242f2d4f691f186d959cbaaa30b0630e4631767b425e" + }, + "stats": { + "downloads": 39677, + "thumbs_up": 1818, + "thumbs_down": 8 + }, + "popularity": { + "downloads": 39677, + "thumbs_up": 1818, + "thumbs_down": 8, + "comments": 89, + "approval": "100% positive (1818/1826)", + "age_days": 6.0, + "downloads_per_day": 522.47, + "rank_by_downloads": 11, + "rank_by_momentum": 2, + "of_total": 150, + "adoption_tier": "Top tier", + "momentum": "Hot · fresh release" + }, + "selection_score": 61893.0, + "selection_reasons": [ + "39,677 downloads", + "1,818 thumbs-up", + "documented triggers + usage notes", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "1248x1824", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "1248x1824", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "1200x1600", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "1200x1600", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "1248x1824", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "1248x1824", + "prompt": "" + } + ], + "tags": [ + "women", + "character", + "instagram", + "realism", + "human", + "model", + "influencer", + "facebook", + "tik tok", + "social media", + "woman", + "sexy" + ], + "scraped_at": "2026-09-05T00:11:58+00:00", + "path": "krea-2/famegrid-krea-2-z-image-qwen-realism-lora" +} diff --git a/granblue-fantasy-style/example_images/01_image.jpeg b/granblue-fantasy-style/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..ed44eed01cbd747ea3afd0da005ab08d8f0cb153 --- /dev/null +++ b/granblue-fantasy-style/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e7b1f2fe17025d75d6b485f966ccd1cdbb753b66091a3d2ed2ece051710c742f +size 8542114 diff --git a/granblue-fantasy-style/example_images/02_image.jpeg b/granblue-fantasy-style/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..cb3c0c19c0def4619dfd448ca3366d0476c68b7f --- /dev/null +++ b/granblue-fantasy-style/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8f40819e8b0156ea8e53bcefcd721850d8d38bef28999b66e450bcdddc9c9c21 +size 1987656 diff --git a/granblue-fantasy-style/example_images/03_image.jpeg b/granblue-fantasy-style/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..7229c4a60e6aa34f20b955609ac3a42aab572a52 --- /dev/null +++ b/granblue-fantasy-style/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e1fa6ed8dfafb29f705ec1637e9477ccbf8d6f760394142907039a7b9c927d4e +size 1495289 diff --git a/granblue-fantasy-style/example_images/04_image.jpeg b/granblue-fantasy-style/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..230c85c74e6cbff53dcde8c78991735aea2f6d01 --- /dev/null +++ b/granblue-fantasy-style/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:81bdf4132e20de74d989a649b6da5a3fd6fd89b0419b6e31319cedf0fd52895f +size 1676964 diff --git a/granblue-fantasy-style/example_images/05_image.jpeg b/granblue-fantasy-style/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..ee31f821aa75dc81de34895ba7bb66d78289bc95 --- /dev/null +++ b/granblue-fantasy-style/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:85284728228a6b70574ca48054be81d567fb28eabbce89f44d76195906ecae40 +size 1293449 diff --git a/granblue-fantasy-style/example_images/06_image.jpeg b/granblue-fantasy-style/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..5d3ef7cf71fefc55912219db584c1a624ebf9760 --- /dev/null +++ b/granblue-fantasy-style/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d2b18eadf4bc7e3eebf2152688b3b41b749c1a98f5abccbe6fa1fdfee6513fbf +size 10983970 diff --git a/granblue-fantasy-style/granblue_1138_detail_l_black_bg_krea2_lora-000008.safetensors b/granblue-fantasy-style/granblue_1138_detail_l_black_bg_krea2_lora-000008.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..a186a6e8e3144e2e38cae94cf3daafe1de332c7f --- /dev/null +++ b/granblue-fantasy-style/granblue_1138_detail_l_black_bg_krea2_lora-000008.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d193fa296df3d306a6d57c4cf5d611d429cd8757a8778a6d9fc1961772ee113f +size 469315488 diff --git a/granblue-fantasy-style/info.txt b/granblue-fantasy-style/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..1650db71336ecb28fab249acf41f33c1844741a3 --- /dev/null +++ b/granblue-fantasy-style/info.txt @@ -0,0 +1,280 @@ +═══════════════════════════════════════════════════════════════════════════════ + GRANBLUE FANTASY STYLE +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Detail / anatomy fixer + Author heathcliff01 + Source https://civitai.com/models/22800 + Provenance CivitAI · model 22800 · version 3090945 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File granblue_1138_detail_l_black_bg_krea2_lora-000008.safetensors + Version granblue for Krea2 + Size 447.6 MB + Format SafeTensor / unknown precision + SHA256 d193fa296df3d306a6d57c4cf5d611d429cd8757a8778a6d9fc1961772ee113f + Published 2026-07-02 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + › "granblue fantasy" + +─────────────────────────────────────────────────────────────────────────────── + RECOMMENDED SETTINGS +─────────────────────────────────────────────────────────────────────────────── + The author did not publish specific settings. + No author settings were published. Community baseline: steps Turbo: 8 + (official README and HF card CLI example; the ComfyUI template ships 8). + Raw: 52 (official README). The krea-2 CLI's generic default is 28. + Community Turbo LoRAs commonly suggest 8-12 (e.g. the Identity Edit card + recommends Turbo at 8-12, or Raw at 20 for removals). A community 4-step + LoRA on Turbo makes 4 valid; its author calls 2 steps an out-of-spec + preview mode., cfg Turbo: CFG disabled - guidance_scale 0.0 in diffusers + and --cfg 0.0 in the krea-2 CLI, which is cfg 1.0 in ComfyUI's KSampler + (the official template ships cfg 1). These are the same thing; do not set + ComfyUI cfg to 0. Raw: CFG 3.5 (official README; CLI generic default 4.5). + Negative prompting on Turbo is effectively unused - the official templates + route ConditioningZeroOut into the negative slot., sampler ComfyUI: + KSampler with sampler euler, scheduler simple, denoise 1.0 (verified in + image_krea2_turbo_t2i.json: [seed, 'randomize', 8, 1, 'euler', 'simple', + 1]). Timestep shift mu = 1.15 for Turbo; ComfyUI bakes this in via + Krea2.sampling_settings = {"multiplier": 1.0, "shift": 1.15}, so the plain + t2i template needs no ModelSampling node, while the style-reference + template adds ModelSamplingFlux at max_shift 1.15 / base_shift 0.5. The CLI + exposes --mu 1.15 for Turbo; Raw uses a resolution-derived mu interpolated + between --y1 0.5 and --y2 1.15 (ai-toolkit mirrors this as base_shift 0.5 / + max_shift 1.15 over image_seq_len 256-6400, exponential time shift)., + resolution 1024x1024 default. Turbo: 1K-2K (the official CLI example + generates 2048x2048; ComfyUI's ResolutionSelector defaults to 1 megapixel, + set 2.0 for 2K). Raw: "trained to generate upto 1k resolution". + Width/height are padded up to a multiple of 16., lora_strength 1.0 for + Krea's nine official style LoRAs (docs.comfy.org lists Recommended Strength + 1.0 for every one), though the shipped t2i template's LoraLoaderModelOnly + is set to 0.8. The style-reference LoRA is used at 1.0 in its template. The + 4-step distill LoRA documents 1.0 as its trained point, 1.0-1.5 as coherent + extrapolation and above ~1.5 as speckle artefacting. Community subject + LoRAs are typically recommended around 0.8-1.2 and dropped toward 0.6-0.9 + when several are stacked.. + +─────────────────────────────────────────────────────────────────────────────── + ADOPTION & POPULARITY +─────────────────────────────────────────────────────────────────────────────── + Downloads 24,647 + Rating 100% positive (2853/2855) + Comments 0 + Published 65 days ago + Download rate 18.6/day since release + + Adoption rank #23 of 150 in this collection (High) + Momentum rank #102 of 150 by download rate (Long tail) + + Adoption tier is the percentile of total downloads within this collection; + momentum is the percentile of downloads-per-day since release. Momentum is + ranked rather than measured against a fixed rate, because this base model + is itself new and every LoRA looks fast on an absolute scale. Both numbers + are CivitAI's, read on the scrape date at the foot of this file. + +─────────────────────────────────────────────────────────────────────────────── + CHAINING / STACKING +─────────────────────────────────────────────────────────────────────────────── + Stacks with others not stated + + The author gave no chaining guidance. The rules below are the + general Krea 2 ones, not anything specific to this LoRA. + + General Krea 2 rules that apply here (see REFERENCE.txt): + + • The most important rule is not about chaining but about which + checkpoint you chain onto: the official README says in bold "TRAIN on + Raw and RUN on Turbo", and "We highly recommend using RAW for training + LoRAs and applying them on Turbo for inference." LoRAs trained on Raw + are stated to work well on Turbo. + • Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 + LoRAs are DiT-only, so there is no CLIP side to chain and the text- + encoder branch stays untouched however many LoRAs you add. + • Krea 2 LoRAs stack in practice and the community treats multi-LoRA as + routine - style + character, style + slider, realism + character. + Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one + writes "I recently stacked over 20 of my own LoRAs on a single + generation and the output still came out great." + • Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge + on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is + loaded. The CivitAI 12GB training guide also notes rank affects + composability: "rank 8 produced a lighter style influence that was + easier to stack with other LoRAs, rank 16 provided a useful balance + between concept strength and composability, and rank 32 was better + suited to a precise object or concept intended to take priority over + other stacked LoRAs." + • Accelerator LoRAs are position-sensitive in intent, not in node order: + a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a + sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to- + Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is + explicitly recommended by the 4-step LoRA's author. + • Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI + (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card + says "Keep mu = 1.15. The training targets are anchored to that grid; a + different shift moves the [sampler off them]"; the same reasoning + applies to any Turbo-trained adapter in the chain. + • Reference/edit LoRAs are designed to be stacked under ordinary subject + and style LoRAs - the Identity Edit card states "Composes with your + LoRAs: character/body/style LoRAs stack on top and steer the prior" and + recommends a stacked subject LoRA when unusual subjects drift toward + the base prior. + • Krea 2 LoRAs cannot be chained with LoRAs for any other base model. + Many CivitAI pages carry multi-model titles like "(Krea-2 + ZIT)" or + "[Flux | ZIB | Krea2]" - those are separate files for separate + architectures shipped under one model page, not a combined adapter. + • To combine two full Krea 2 checkpoints rather than adapters, ComfyUI + ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per- + section weights for first., tmlp., txtmlp., tproj., + txtfusion.projector, txtfusion.layerwise_blocks.0-1, + txtfusion.refiner_blocks.0-1, blocks.0-27 and last. + • The author never states which weight family this targets. Silent + degradation, never a hard error, between Raw and Turbo. The two share + one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" + - so every LoRA tensor shape matches on both and ComfyUI/diffuser + +─────────────────────────────────────────────────────────────────────────────── + PROS +─────────────────────────────────────────────────────────────────────────────── + + Widely adopted — 24,647 downloads. [source: CivitAI stats] + + Positively received — 2,853 thumbs-up, 2 down. [source: CivitAI stats] + + Explicit trigger word(s) — granblue fantasy — so the effect can be + turned on and off from the prompt. [source: model metadata] + +─────────────────────────────────────────────────────────────────────────────── + CONS & CAVEATS +─────────────────────────────────────────────────────────────────────────────── + - Author never states which generation mode it was trained for; test + against your own workflow before relying on it. [source: description] + +─────────────────────────────────────────────────────────────────────────────── + COMPATIBILITY WARNING +─────────────────────────────────────────────────────────────────────────────── + Silent degradation, never a hard error, between Raw and Turbo. The two + share one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" - so + every LoRA tensor shape matches on both and ComfyUI/diffusers will load an + adapter against either without complaint. The lvladikov 4-step LoRA card + states this for the reverse direction: "Every layer it targets also exists + in Krea 2 Raw, so it will load there without complaint - but that is a side + effect of the shared architecture, not a supported mode", and reports + mixed, subject-dependent results with malformed subjects (duplicated heads, + fused limbs) at low step counts, needing 14+ steps with Raw's CFG on before + output is coherent. The intended direction (train on Raw, run on Turbo) is + officially blessed. The more common practical failure is a settings + mismatch rather than a weight mismatch: running a Turbo-targeted LoRA at + Raw's 52 steps + CFG 3.5, or leaving CFG enabled on Turbo, degrades output + with no warning. Loading a Krea 2 LoRA against a different model family + entirely (Flux, Qwen-Image, Z-Image) is also not an error - ComfyUI logs + "lora key not loaded: {key}" per key and generates as if no LoRA were + applied. + +─────────────────────────────────────────────────────────────────────────────── + AUTHOR'S DESCRIPTION +─────────────────────────────────────────────────────────────────────────────── + Only initial vertical drawing is used as the training set, and breakthrough + vertical drawing is not used.So there aren't many additional effects. + For newbie image version, please git clone https://github.com/NewBieAI- + Lab/ComfyUI-Newbie-Nodes.git + and use comfy_newbie_lora_loader + to load this lora. + +─────────────────────────────────────────────────────────────────────────────── + VERSION NOTES +─────────────────────────────────────────────────────────────────────────────── + this version using 800+ images and trained with tag and nl. + +─────────────────────────────────────────────────────────────────────────────── + EXAMPLE MEDIA +─────────────────────────────────────────────────────────────────────────────── + 6 preview(s) in ./example_images/ — community results for + this LoRA. Videos keep a matching _poster.jpg still frame. + + No prompts are recorded below: CivitAI's public API flags these posts as + having generation metadata but does not return it, so the prompts cannot be + scraped. Open the model page to read them. + + 01_image.jpeg (image, 2048x3072) + 02_image.jpeg (image, 2048x3072) + 03_image.jpeg (image, 1024x1024) + 04_image.jpeg (image, 1024x1024) + 05_image.jpeg (image, 840x1256) + 06_image.jpeg (image, 2560x3680) + +─────────────────────────────────────────────────────────────────────────────── + HOW TO USE (COMFYUI) +─────────────────────────────────────────────────────────────────────────────── + 1. Update ComfyUI to at least v0.26.0 (released 2026-06-23), which added + local Krea2 support. For reference-image / style-reference / edit LoRAs you + need at least v0.29.0 (2026-07-29), which added regular and timestep-zero + reference images for Krea 2. ModelMergeKrea2 arrived in v0.27.0 + (2026-06-30). + + 2. Put the base stack in place (files from huggingface.co/Comfy-Org/Krea-2, + which is not gated): + - ComfyUI/models/diffusion_models/krea2_turbo_fp8_scaled.safetensors + (documented recommended default; the repo also holds krea2_turbo_bf16, + _int8_convrot, _mxfp8, _nvfp4 and krea2_raw_bf16 / _fp8_scaled / + _int8_convrot) + - ComfyUI/models/text_encoders/qwen3vl_4b_fp8_scaled.safetensors (bf16 + also available) + - ComfyUI/models/vae/qwen_image_vae.safetensors + + 3. Put the LoRA .safetensors in ComfyUI/models/loras/ - flat, no per-model + subfolder required. + + 4. Build or load the graph. The official image_krea2_turbo_t2i.json + template is: UNETLoader (krea2_turbo_fp8_scaled.safetensors) -> + LoraLoaderModelOnly -> KSampler; CLIPLoader + (qwen3vl_4b_fp8_scaled.safetensors, type "krea2") -> text encode -> + KSampler positive, with ConditioningZeroOut into negative; VAELoader + (qwen_image_vae.safetensors) -> VAEDecode -> SaveImage; EmptyLatentImage + 1024x1024 driven by ResolutionSelector. An int8 sibling template + (image_krea2_turbo_t2i_int8.json) also ships. + + 5. Load the LoRA with LoraLoaderModelOnly, not LoraLoader - both official + templates do. Krea 2 LoRAs patch the DiT only; note that the DiT's own + text_fusion / txtfusion blocks are a normal LoRA target (the official style + LoRAs patch them), but the Qwen3-VL text encoder itself is not. Set + strength 1.0 for Krea's official style LoRAs (the shipped template uses + 0.8) and add the LoRA's trigger word to the prompt. + + 6. Sampler settings: steps 8, cfg 1, sampler euler, scheduler simple, + denoise 1.0. Do not add a ModelSampling node for plain t2i - ComfyUI + already applies shift 1.15 from Krea2.sampling_settings. + + 7. For style reference, use the dedicated template: UNETLoader + krea2_turbo_int8_convrot.safetensors, LoraLoaderModelOnly + krea2_style_reference.safetensors at 1.0, ModelSamplingFlux (1.15, 0.5), + TextEncodeQwenImageEditPlus fed by LoadImage, + FluxKontextMultiReferenceLatentMethod set to index_timestep_zero, and + SamplerCustomAdvanced with KSamplerSelect euler + BasicScheduler simple/8 + + CFGGuider cfg 1. + + 8. For instruction editing (Krea 2 Identity Edit), stock nodes are not + enough: install the ComfyUI-Krea2Edit node pack, which ships its own + workflows, and run Turbo at 8-12 steps / cfg 1 (or Raw at 20 steps / cfg 3 + for removals) with LoRA strength 1.0. + + 9. Low-VRAM alternative: install city96's ComfyUI-GGUF and use a community + GGUF build (realrebelai/KREA-2_GGUFs, quantized from Turbo; + molbal/krea2-gguf, quantized from Raw) with the Unet Loader (GGUF) node in + place of UNETLoader. There is no official GGUF release. LoRA loading is + unchanged. + +─────────────────────────────────────────────────────────────────────────────── + Scraped by lorakit on 2026-09-05 00:11 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/granblue-fantasy-style/metadata.json b/granblue-fantasy-style/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..6696cc6bcad5c4f82ec111c8c76c7e69388acecc --- /dev/null +++ b/granblue-fantasy-style/metadata.json @@ -0,0 +1,131 @@ +{ + "name": "Granblue fantasy style", + "slug": "granblue-fantasy-style", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 22800, + "version_id": 3090945, + "version_name": "granblue for Krea2", + "source": "civitai", + "url": "https://civitai.com/models/22800", + "repo_id": null, + "file_path": null, + "author": "heathcliff01", + "category": { + "key": "detail", + "label": "Detail / anatomy fixer" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [ + "granblue fantasy" + ], + "recommended_settings": {}, + "chaining": { + "author": { + "stacks": "not stated", + "notes": "", + "avoid": [], + "requires": [], + "order": "", + "strength": "", + "stated": false + }, + "general_rules": [ + "The most important rule is not about chaining but about which checkpoint you chain onto: the official README says in bold \"TRAIN on Raw and RUN on Turbo\", and \"We highly recommend using RAW for training LoRAs and applying them on Turbo for inference.\" LoRAs trained on Raw are stated to work well on Turbo.", + "Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 LoRAs are DiT-only, so there is no CLIP side to chain and the text-encoder branch stays untouched however many LoRAs you add.", + "Krea 2 LoRAs stack in practice and the community treats multi-LoRA as routine - style + character, style + slider, realism + character. Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one writes \"I recently stacked over 20 of my own LoRAs on a single generation and the output still came out great.\"", + "Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is loaded. The CivitAI 12GB training guide also notes rank affects composability: \"rank 8 produced a lighter style influence that was easier to stack with other LoRAs, rank 16 provided a useful balance between concept strength and composability, and rank 32 was better suited to a precise object or concept intended to take priority over other stacked LoRAs.\"", + "Accelerator LoRAs are position-sensitive in intent, not in node order: a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to-Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is explicitly recommended by the 4-step LoRA's author.", + "Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card says \"Keep mu = 1.15. The training targets are anchored to that grid; a different shift moves the [sampler off them]\"; the same reasoning applies to any Turbo-trained adapter in the chain.", + "Reference/edit LoRAs are designed to be stacked under ordinary subject and style LoRAs - the Identity Edit card states \"Composes with your LoRAs: character/body/style LoRAs stack on top and steer the prior\" and recommends a stacked subject LoRA when unusual subjects drift toward the base prior.", + "Krea 2 LoRAs cannot be chained with LoRAs for any other base model. Many CivitAI pages carry multi-model titles like \"(Krea-2 + ZIT)\" or \"[Flux | ZIB | Krea2]\" - those are separate files for separate architectures shipped under one model page, not a combined adapter.", + "To combine two full Krea 2 checkpoints rather than adapters, ComfyUI ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per-section weights for first., tmlp., txtmlp., tproj., txtfusion.projector, txtfusion.layerwise_blocks.0-1, txtfusion.refiner_blocks.0-1, blocks.0-27 and last.", + "The author never states which weight family this targets. Silent degradation, never a hard error, between Raw and Turbo. The two share one architecture - ai-toolkit's config is commented \"The reference 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)\" - so every LoRA tensor shape matches on both and ComfyUI/diffuser" + ] + }, + "pros": [ + "Widely adopted — 24,647 downloads. [source: CivitAI stats]", + "Positively received — 2,853 thumbs-up, 2 down. [source: CivitAI stats]", + "Explicit trigger word(s) — granblue fantasy — so the effect can be turned on and off from the prompt. [source: model metadata]" + ], + "cons": [ + "Author never states which generation mode it was trained for; test against your own workflow before relying on it. [source: description]" + ], + "file": { + "name": "granblue_1138_detail_l_black_bg_krea2_lora-000008.safetensors", + "size_bytes": 469315488, + "size_mb": 447.57, + "sha256": "d193fa296df3d306a6d57c4cf5d611d429cd8757a8778a6d9fc1961772ee113f" + }, + "stats": { + "downloads": 24647, + "thumbs_up": 2853, + "thumbs_down": 2 + }, + "popularity": { + "downloads": 24647, + "thumbs_up": 2853, + "thumbs_down": 2, + "comments": 0, + "approval": "100% positive (2853/2855)", + "age_days": 64.6, + "downloads_per_day": 18.56, + "rank_by_downloads": 23, + "rank_by_momentum": 102, + "of_total": 150, + "adoption_tier": "High", + "momentum": "Long tail" + }, + "selection_score": 58883.0, + "selection_reasons": [ + "24,647 downloads", + "2,853 thumbs-up", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "2048x3072", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "2048x3072", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "1024x1024", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "1024x1024", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "840x1256", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "2560x3680", + "prompt": "" + } + ], + "tags": [ + "style", + "game character" + ], + "scraped_at": "2026-09-05T00:11:56+00:00", + "path": "krea-2/granblue-fantasy-style" +} diff --git a/krea2-textfusion-refusal-reduction-lora/Krea2_TextFusion_Refusal_Reduction.safetensors b/krea2-textfusion-refusal-reduction-lora/Krea2_TextFusion_Refusal_Reduction.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..e2508286f2d548f82e34064c2ceb59f2445d004e --- /dev/null +++ b/krea2-textfusion-refusal-reduction-lora/Krea2_TextFusion_Refusal_Reduction.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:84ec722ddab93f6489c5315bca25de5dd1a7b7ec5045a3c4ce2f97f62e54e8e6 +size 27665664 diff --git a/krea2-textfusion-refusal-reduction-lora/example_images/01_image.jpeg b/krea2-textfusion-refusal-reduction-lora/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..fbfaf58e8a9ab9a296621730c3d1b34ad5bcb740 --- /dev/null +++ b/krea2-textfusion-refusal-reduction-lora/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a351cb52d74647953b419dab2a297f0f58166e6d1dd1228ac57a8613cb3d2695 +size 381051 diff --git a/krea2-textfusion-refusal-reduction-lora/example_images/02_image.jpeg b/krea2-textfusion-refusal-reduction-lora/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..25e96b196d31dcf21e15740f31c7ab8c4ec6cd68 --- /dev/null +++ b/krea2-textfusion-refusal-reduction-lora/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e1e6862275f0a76b674fe504d89f74d4f7429c498826e45c5f67086e76bdb4c +size 214454 diff --git a/krea2-textfusion-refusal-reduction-lora/example_images/03_image.jpeg b/krea2-textfusion-refusal-reduction-lora/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..b819213672912e5199685bad901a62d3f34811d5 --- /dev/null +++ b/krea2-textfusion-refusal-reduction-lora/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fbacbc977460b930e63eccf9f82d9bedf47f8c920869cb45db6678b44b112cc3 +size 1217174 diff --git a/krea2-textfusion-refusal-reduction-lora/example_images/04_image.jpeg b/krea2-textfusion-refusal-reduction-lora/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..5b3e59878cfe8b2bbb8b9a4819ad258a712a8938 --- /dev/null +++ b/krea2-textfusion-refusal-reduction-lora/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:db54ad58b3ce60d2572d26efb070d4b72545f8f97d965e517788ab698cd1a3ed +size 1851509 diff --git a/krea2-textfusion-refusal-reduction-lora/example_images/05_image.jpeg b/krea2-textfusion-refusal-reduction-lora/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..ff3fa22438774c6c2b3c54a5438af28bbb4ea3ce --- /dev/null +++ b/krea2-textfusion-refusal-reduction-lora/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6a154da2c63c6c7ffb7a2c5f9cfcb522d074b08f7f8c7b7202e7622f529a009e +size 246922 diff --git a/krea2-textfusion-refusal-reduction-lora/example_images/06_image.jpeg b/krea2-textfusion-refusal-reduction-lora/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..8dab9020882023e0a115dd9e46160fbd387ce231 --- /dev/null +++ b/krea2-textfusion-refusal-reduction-lora/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:35208d78c4a7ee462506f9a75fbb667c85705df31d50874442dc12e940073bb7 +size 852506 diff --git a/krea2-textfusion-refusal-reduction-lora/info.txt b/krea2-textfusion-refusal-reduction-lora/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..295c43ca0f3e628179b99fb19a4938170d3db184 --- /dev/null +++ b/krea2-textfusion-refusal-reduction-lora/info.txt @@ -0,0 +1,340 @@ +═══════════════════════════════════════════════════════════════════════════════ + KREA2 TEXTFUSION REFUSAL-REDUCTION LORA +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Detail / anatomy fixer + Author Capitan01R + Source https://civitai.com/models/2775340 + Provenance CivitAI · model 2775340 · version 3125118 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File Krea2_TextFusion_Refusal_Reduction.safetensors + Version v1.0 + Size 26.4 MB + Format SafeTensor / unknown precision + SHA256 84ec722ddab93f6489c5315bca25de5dd1a7b7ec5045a3c4ce2f97f62e54e8e6 + Published 2026-07-13 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + No trigger word. The LoRA applies as soon as it is loaded — + strength is the only way to dial the effect up or down. + +─────────────────────────────────────────────────────────────────────────────── + RECOMMENDED SETTINGS +─────────────────────────────────────────────────────────────────────────────── + The author did not publish specific settings. + No author settings were published. Community baseline: steps Turbo: 8 + (official README and HF card CLI example; the ComfyUI template ships 8). + Raw: 52 (official README). The krea-2 CLI's generic default is 28. + Community Turbo LoRAs commonly suggest 8-12 (e.g. the Identity Edit card + recommends Turbo at 8-12, or Raw at 20 for removals). A community 4-step + LoRA on Turbo makes 4 valid; its author calls 2 steps an out-of-spec + preview mode., cfg Turbo: CFG disabled - guidance_scale 0.0 in diffusers + and --cfg 0.0 in the krea-2 CLI, which is cfg 1.0 in ComfyUI's KSampler + (the official template ships cfg 1). These are the same thing; do not set + ComfyUI cfg to 0. Raw: CFG 3.5 (official README; CLI generic default 4.5). + Negative prompting on Turbo is effectively unused - the official templates + route ConditioningZeroOut into the negative slot., sampler ComfyUI: + KSampler with sampler euler, scheduler simple, denoise 1.0 (verified in + image_krea2_turbo_t2i.json: [seed, 'randomize', 8, 1, 'euler', 'simple', + 1]). Timestep shift mu = 1.15 for Turbo; ComfyUI bakes this in via + Krea2.sampling_settings = {"multiplier": 1.0, "shift": 1.15}, so the plain + t2i template needs no ModelSampling node, while the style-reference + template adds ModelSamplingFlux at max_shift 1.15 / base_shift 0.5. The CLI + exposes --mu 1.15 for Turbo; Raw uses a resolution-derived mu interpolated + between --y1 0.5 and --y2 1.15 (ai-toolkit mirrors this as base_shift 0.5 / + max_shift 1.15 over image_seq_len 256-6400, exponential time shift)., + resolution 1024x1024 default. Turbo: 1K-2K (the official CLI example + generates 2048x2048; ComfyUI's ResolutionSelector defaults to 1 megapixel, + set 2.0 for 2K). Raw: "trained to generate upto 1k resolution". + Width/height are padded up to a multiple of 16., lora_strength 1.0 for + Krea's nine official style LoRAs (docs.comfy.org lists Recommended Strength + 1.0 for every one), though the shipped t2i template's LoraLoaderModelOnly + is set to 0.8. The style-reference LoRA is used at 1.0 in its template. The + 4-step distill LoRA documents 1.0 as its trained point, 1.0-1.5 as coherent + extrapolation and above ~1.5 as speckle artefacting. Community subject + LoRAs are typically recommended around 0.8-1.2 and dropped toward 0.6-0.9 + when several are stacked.. + +─────────────────────────────────────────────────────────────────────────────── + ADOPTION & POPULARITY +─────────────────────────────────────────────────────────────────────────────── + Downloads 29,758 + Rating 100% positive (1702/1702) + Comments 0 + Published 54 days ago + Download rate 553.8/day since release + + Adoption rank #19 of 150 in this collection (High) + Momentum rank #1 of 150 by download rate (Hot) + + Adoption tier is the percentile of total downloads within this collection; + momentum is the percentile of downloads-per-day since release. Momentum is + ranked rather than measured against a fixed rate, because this base model + is itself new and every LoRA looks fast on an absolute scale. Both numbers + are CivitAI's, read on the scrape date at the foot of this file. + +─────────────────────────────────────────────────────────────────────────────── + CHAINING / STACKING +─────────────────────────────────────────────────────────────────────────────── + Stacks with others not stated + + The author gave no chaining guidance. The rules below are the + general Krea 2 ones, not anything specific to this LoRA. + + General Krea 2 rules that apply here (see REFERENCE.txt): + + • The most important rule is not about chaining but about which + checkpoint you chain onto: the official README says in bold "TRAIN on + Raw and RUN on Turbo", and "We highly recommend using RAW for training + LoRAs and applying them on Turbo for inference." LoRAs trained on Raw + are stated to work well on Turbo. + • Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 + LoRAs are DiT-only, so there is no CLIP side to chain and the text- + encoder branch stays untouched however many LoRAs you add. + • Krea 2 LoRAs stack in practice and the community treats multi-LoRA as + routine - style + character, style + slider, realism + character. + Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one + writes "I recently stacked over 20 of my own LoRAs on a single + generation and the output still came out great." + • Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge + on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is + loaded. The CivitAI 12GB training guide also notes rank affects + composability: "rank 8 produced a lighter style influence that was + easier to stack with other LoRAs, rank 16 provided a useful balance + between concept strength and composability, and rank 32 was better + suited to a precise object or concept intended to take priority over + other stacked LoRAs." + • Accelerator LoRAs are position-sensitive in intent, not in node order: + a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a + sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to- + Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is + explicitly recommended by the 4-step LoRA's author. + • Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI + (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card + says "Keep mu = 1.15. The training targets are anchored to that grid; a + different shift moves the [sampler off them]"; the same reasoning + applies to any Turbo-trained adapter in the chain. + • Reference/edit LoRAs are designed to be stacked under ordinary subject + and style LoRAs - the Identity Edit card states "Composes with your + LoRAs: character/body/style LoRAs stack on top and steer the prior" and + recommends a stacked subject LoRA when unusual subjects drift toward + the base prior. + • Krea 2 LoRAs cannot be chained with LoRAs for any other base model. + Many CivitAI pages carry multi-model titles like "(Krea-2 + ZIT)" or + "[Flux | ZIB | Krea2]" - those are separate files for separate + architectures shipped under one model page, not a combined adapter. + • To combine two full Krea 2 checkpoints rather than adapters, ComfyUI + ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per- + section weights for first., tmlp., txtmlp., tproj., + txtfusion.projector, txtfusion.layerwise_blocks.0-1, + txtfusion.refiner_blocks.0-1, blocks.0-27 and last. + • The author never states which weight family this targets. Silent + degradation, never a hard error, between Raw and Turbo. The two share + one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" + - so every LoRA tensor shape matches on both and ComfyUI/diffuser + +─────────────────────────────────────────────────────────────────────────────── + PROS +─────────────────────────────────────────────────────────────────────────────── + + Widely adopted — 29,758 downloads. [source: CivitAI stats] + + Positively received — 1,702 thumbs-up with no down-votes. [source: + CivitAI stats] + + Compact at 26 MB — cheap to stack with other LoRAs. [source: file size] + +─────────────────────────────────────────────────────────────────────────────── + CONS & CAVEATS +─────────────────────────────────────────────────────────────────────────────── + - No trigger word — the LoRA is always on once loaded, so strength is the + only control. [source: model metadata] + - Author never states which generation mode it was trained for; test + against your own workflow before relying on it. [source: description] + +─────────────────────────────────────────────────────────────────────────────── + COMPATIBILITY WARNING +─────────────────────────────────────────────────────────────────────────────── + Silent degradation, never a hard error, between Raw and Turbo. The two + share one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" - so + every LoRA tensor shape matches on both and ComfyUI/diffusers will load an + adapter against either without complaint. The lvladikov 4-step LoRA card + states this for the reverse direction: "Every layer it targets also exists + in Krea 2 Raw, so it will load there without complaint - but that is a side + effect of the shared architecture, not a supported mode", and reports + mixed, subject-dependent results with malformed subjects (duplicated heads, + fused limbs) at low step counts, needing 14+ steps with Raw's CFG on before + output is coherent. The intended direction (train on Raw, run on Turbo) is + officially blessed. The more common practical failure is a settings + mismatch rather than a weight mismatch: running a Turbo-targeted LoRA at + Raw's 52 steps + CFG 3.5, or leaving CFG enabled on Turbo, degrades output + with no warning. Loading a Krea 2 LoRA against a different model family + entirely (Flux, Qwen-Image, Z-Image) is also not an error - ComfyUI logs + "lora key not loaded: {key}" per key and generates as if no LoRA were + applied. + +─────────────────────────────────────────────────────────────────────────────── + AUTHOR'S DESCRIPTION +─────────────────────────────────────────────────────────────────────────────── + Progress update, + the model refusal behaviour is deeper than the textfusion path, Dit blocks + are included in this behaviour; therefore a LoRA alone CANNOT fix this + issue, I am going to be working on a custom node that targets the DiT + blocks without wrecking the model. + + V1.0 I recommend loading this LoRA with this LoRA Loader I just created : + https://civitai.com/models/2844700/comfyui-sigmasync-lora + Buy me a coffee :) + + # Krea2 TextFusion Refusal-Reduction LoRA + This is a dedicated rank-64 LoRA trained with one narrow objective: reduce + Krea2’s learned refusal and restriction behavior while preserving as much + of the base model’s existing visual knowledge and prompt behavior as + possible. + This is not a concept, character, style, anatomy, or aesthetic LoRA. It was + not trained to add new visual concepts, reproduce a training-image set, or + impose a preferred look. Its training objective was exclusively to push the + model away from refusal behavior so that concepts already represented + within the base model are less likely to be suppressed during text + conditioning. + ## What It Changes + Krea2 receives multiple hidden-state taps from its Qwen-VL text encoder and + processes them through TextFusion before sending the resulting conditioning + into the image transformer. This LoRA applies learned low-rank residuals + only to the attention and internal MLP projections within: + - txtfusion.layerwise_blocks.0 + - txtfusion.layerwise_blocks.1 + - txtfusion.refiner_blocks.0 + - txtfusion.refiner_blocks.1 + The release version deliberately contains no adapter for: + - The TextFusion 1 × 12 tap projector + - The external/general txtmlp + - The image-transformer blocks or any other image-generation layers + ## How It Works + The base checkpoint is never overwritten. For every targeted linear layer, + LoRA adds a learned rank-64 residual to the original weight during + inference: + W_effective = W_base + strength × ΔW + This changes how existing Qwen-VL text features are routed, gated, + transported, and refined through TextFusion before they reach Krea2’s + untouched image transformer. In other words, the LoRA is intended to + improve access to visual knowledge already present in the base model, not + inject a new concept or replace the model’s learned visual representations. + This release targets the specific TextFusion route isolated through layer- + by-layer ablation instead of broadly amplifying activations or directly + altering the projector. + ## Usage + Use strength of "1.00" + + Responsibility and Output Disclaimer + This LoRA modifies the model’s text-conditioning behavior but does not + determine, authorize, endorse, supervise, or control the images produced by + individual users, prompts, workflows, checkpoints, samplers, or third-party + software. + Generated outputs remain dependent on the base model, user-provided + conditioning, inference configuration, and the surrounding generation + pipeline. The author assumes no responsibility or liability for what users + choose to generate, how generated material is used, or whether any output + complies with applicable laws, platform rules, licensing terms, or third- + party rights. + Users are solely responsible for their prompts, generated outputs, + distribution decisions, and use of this LoRA. + +─────────────────────────────────────────────────────────────────────────────── + VERSION NOTES +─────────────────────────────────────────────────────────────────────────────── + v.1.0.0 Krea 2 Turbo + +─────────────────────────────────────────────────────────────────────────────── + EXAMPLE MEDIA +─────────────────────────────────────────────────────────────────────────────── + 6 preview(s) in ./example_images/ — community results for + this LoRA. Videos keep a matching _poster.jpg still frame. + + No prompts are recorded below: CivitAI's public API flags these posts as + having generation metadata but does not return it, so the prompts cannot be + scraped. Open the model page to read them. + + 01_image.jpeg (image, 1024x1024) + 02_image.jpeg (image, 832x1248) + 03_image.jpeg (image, 1664x2432) + 04_image.jpeg (image, 1664x2432) + 05_image.jpeg (image, 832x1248) + 06_image.jpeg (image, 1664x2432) + +─────────────────────────────────────────────────────────────────────────────── + HOW TO USE (COMFYUI) +─────────────────────────────────────────────────────────────────────────────── + 1. Update ComfyUI to at least v0.26.0 (released 2026-06-23), which added + local Krea2 support. For reference-image / style-reference / edit LoRAs you + need at least v0.29.0 (2026-07-29), which added regular and timestep-zero + reference images for Krea 2. ModelMergeKrea2 arrived in v0.27.0 + (2026-06-30). + + 2. Put the base stack in place (files from huggingface.co/Comfy-Org/Krea-2, + which is not gated): + - ComfyUI/models/diffusion_models/krea2_turbo_fp8_scaled.safetensors + (documented recommended default; the repo also holds krea2_turbo_bf16, + _int8_convrot, _mxfp8, _nvfp4 and krea2_raw_bf16 / _fp8_scaled / + _int8_convrot) + - ComfyUI/models/text_encoders/qwen3vl_4b_fp8_scaled.safetensors (bf16 + also available) + - ComfyUI/models/vae/qwen_image_vae.safetensors + + 3. Put the LoRA .safetensors in ComfyUI/models/loras/ - flat, no per-model + subfolder required. + + 4. Build or load the graph. The official image_krea2_turbo_t2i.json + template is: UNETLoader (krea2_turbo_fp8_scaled.safetensors) -> + LoraLoaderModelOnly -> KSampler; CLIPLoader + (qwen3vl_4b_fp8_scaled.safetensors, type "krea2") -> text encode -> + KSampler positive, with ConditioningZeroOut into negative; VAELoader + (qwen_image_vae.safetensors) -> VAEDecode -> SaveImage; EmptyLatentImage + 1024x1024 driven by ResolutionSelector. An int8 sibling template + (image_krea2_turbo_t2i_int8.json) also ships. + + 5. Load the LoRA with LoraLoaderModelOnly, not LoraLoader - both official + templates do. Krea 2 LoRAs patch the DiT only; note that the DiT's own + text_fusion / txtfusion blocks are a normal LoRA target (the official style + LoRAs patch them), but the Qwen3-VL text encoder itself is not. Set + strength 1.0 for Krea's official style LoRAs (the shipped template uses + 0.8) and add the LoRA's trigger word to the prompt. + + 6. Sampler settings: steps 8, cfg 1, sampler euler, scheduler simple, + denoise 1.0. Do not add a ModelSampling node for plain t2i - ComfyUI + already applies shift 1.15 from Krea2.sampling_settings. + + 7. For style reference, use the dedicated template: UNETLoader + krea2_turbo_int8_convrot.safetensors, LoraLoaderModelOnly + krea2_style_reference.safetensors at 1.0, ModelSamplingFlux (1.15, 0.5), + TextEncodeQwenImageEditPlus fed by LoadImage, + FluxKontextMultiReferenceLatentMethod set to index_timestep_zero, and + SamplerCustomAdvanced with KSamplerSelect euler + BasicScheduler simple/8 + + CFGGuider cfg 1. + + 8. For instruction editing (Krea 2 Identity Edit), stock nodes are not + enough: install the ComfyUI-Krea2Edit node pack, which ships its own + workflows, and run Turbo at 8-12 steps / cfg 1 (or Raw at 20 steps / cfg 3 + for removals) with LoRA strength 1.0. + + 9. Low-VRAM alternative: install city96's ComfyUI-GGUF and use a community + GGUF build (realrebelai/KREA-2_GGUFs, quantized from Turbo; + molbal/krea2-gguf, quantized from Raw) with the Unet Loader (GGUF) node in + place of UNETLoader. There is no official GGUF release. LoRA loading is + unchanged. + +─────────────────────────────────────────────────────────────────────────────── + Scraped by lorakit on 2026-09-05 00:12 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/krea2-textfusion-refusal-reduction-lora/metadata.json b/krea2-textfusion-refusal-reduction-lora/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..862d8a739fe25b53f9200d885917131fcfa950d2 --- /dev/null +++ b/krea2-textfusion-refusal-reduction-lora/metadata.json @@ -0,0 +1,130 @@ +{ + "name": "Krea2 TextFusion Refusal-Reduction LoRA", + "slug": "krea2-textfusion-refusal-reduction-lora", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 2775340, + "version_id": 3125118, + "version_name": "v1.0", + "source": "civitai", + "url": "https://civitai.com/models/2775340", + "repo_id": null, + "file_path": null, + "author": "Capitan01R", + "category": { + "key": "detail", + "label": "Detail / anatomy fixer" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [], + "recommended_settings": {}, + "chaining": { + "author": { + "stacks": "not stated", + "notes": "", + "avoid": [], + "requires": [], + "order": "", + "strength": "", + "stated": false + }, + "general_rules": [ + "The most important rule is not about chaining but about which checkpoint you chain onto: the official README says in bold \"TRAIN on Raw and RUN on Turbo\", and \"We highly recommend using RAW for training LoRAs and applying them on Turbo for inference.\" LoRAs trained on Raw are stated to work well on Turbo.", + "Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 LoRAs are DiT-only, so there is no CLIP side to chain and the text-encoder branch stays untouched however many LoRAs you add.", + "Krea 2 LoRAs stack in practice and the community treats multi-LoRA as routine - style + character, style + slider, realism + character. Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one writes \"I recently stacked over 20 of my own LoRAs on a single generation and the output still came out great.\"", + "Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is loaded. The CivitAI 12GB training guide also notes rank affects composability: \"rank 8 produced a lighter style influence that was easier to stack with other LoRAs, rank 16 provided a useful balance between concept strength and composability, and rank 32 was better suited to a precise object or concept intended to take priority over other stacked LoRAs.\"", + "Accelerator LoRAs are position-sensitive in intent, not in node order: a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to-Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is explicitly recommended by the 4-step LoRA's author.", + "Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card says \"Keep mu = 1.15. The training targets are anchored to that grid; a different shift moves the [sampler off them]\"; the same reasoning applies to any Turbo-trained adapter in the chain.", + "Reference/edit LoRAs are designed to be stacked under ordinary subject and style LoRAs - the Identity Edit card states \"Composes with your LoRAs: character/body/style LoRAs stack on top and steer the prior\" and recommends a stacked subject LoRA when unusual subjects drift toward the base prior.", + "Krea 2 LoRAs cannot be chained with LoRAs for any other base model. Many CivitAI pages carry multi-model titles like \"(Krea-2 + ZIT)\" or \"[Flux | ZIB | Krea2]\" - those are separate files for separate architectures shipped under one model page, not a combined adapter.", + "To combine two full Krea 2 checkpoints rather than adapters, ComfyUI ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per-section weights for first., tmlp., txtmlp., tproj., txtfusion.projector, txtfusion.layerwise_blocks.0-1, txtfusion.refiner_blocks.0-1, blocks.0-27 and last.", + "The author never states which weight family this targets. Silent degradation, never a hard error, between Raw and Turbo. The two share one architecture - ai-toolkit's config is commented \"The reference 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)\" - so every LoRA tensor shape matches on both and ComfyUI/diffuser" + ] + }, + "pros": [ + "Widely adopted — 29,758 downloads. [source: CivitAI stats]", + "Positively received — 1,702 thumbs-up with no down-votes. [source: CivitAI stats]", + "Compact at 26 MB — cheap to stack with other LoRAs. [source: file size]" + ], + "cons": [ + "No trigger word — the LoRA is always on once loaded, so strength is the only control. [source: model metadata]", + "Author never states which generation mode it was trained for; test against your own workflow before relying on it. [source: description]" + ], + "file": { + "name": "Krea2_TextFusion_Refusal_Reduction.safetensors", + "size_bytes": 27665664, + "size_mb": 26.38, + "sha256": "84ec722ddab93f6489c5315bca25de5dd1a7b7ec5045a3c4ce2f97f62e54e8e6" + }, + "stats": { + "downloads": 29758, + "thumbs_up": 1702, + "thumbs_down": 0 + }, + "popularity": { + "downloads": 29758, + "thumbs_up": 1702, + "thumbs_down": 0, + "comments": 0, + "approval": "100% positive (1702/1702)", + "age_days": 53.7, + "downloads_per_day": 553.85, + "rank_by_downloads": 19, + "rank_by_momentum": 1, + "of_total": 150, + "adoption_tier": "High", + "momentum": "Hot" + }, + "selection_score": 50182.0, + "selection_reasons": [ + "29,758 downloads", + "1,702 thumbs-up", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "1024x1024", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "832x1248", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "1664x2432", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "1664x2432", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "832x1248", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "1664x2432", + "prompt": "" + } + ], + "tags": [ + "krea 2 turbo", + "tool" + ], + "scraped_at": "2026-09-05T00:12:02+00:00", + "path": "krea-2/krea2-textfusion-refusal-reduction-lora" +} diff --git a/new-fantasy-core-ill-flux-pony-sdxl-zit-detailer/NewFantasyCore_Krea2.safetensors b/new-fantasy-core-ill-flux-pony-sdxl-zit-detailer/NewFantasyCore_Krea2.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..fa0714def6b441d578aebb6a470dad6823608123 --- /dev/null +++ b/new-fantasy-core-ill-flux-pony-sdxl-zit-detailer/NewFantasyCore_Krea2.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eab3cd9df8b5673cc6f5895bfbf4fd3814eb0df4d3fe2b6abd1b43d2398ee374 +size 228588904 diff --git a/new-fantasy-core-ill-flux-pony-sdxl-zit-detailer/example_images/01_image.jpeg b/new-fantasy-core-ill-flux-pony-sdxl-zit-detailer/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..8d9a930dba4a58b1ad05c851a7e42943ee952338 --- /dev/null +++ b/new-fantasy-core-ill-flux-pony-sdxl-zit-detailer/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:74b4ae6cd0b4845dcf5da602c9e6ca860eb7ca08eea7f0b580f65b41d614d999 +size 2309401 diff --git a/new-fantasy-core-ill-flux-pony-sdxl-zit-detailer/example_images/02_image.jpeg b/new-fantasy-core-ill-flux-pony-sdxl-zit-detailer/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..2ab37baf8fac9639147e6b71ec063cfc8e370a2b --- /dev/null +++ b/new-fantasy-core-ill-flux-pony-sdxl-zit-detailer/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6510489a560c2d02f942910f6d409938e7e853462239fea941d460fe915c1f73 +size 2166423 diff --git a/new-fantasy-core-ill-flux-pony-sdxl-zit-detailer/example_images/03_image.jpeg b/new-fantasy-core-ill-flux-pony-sdxl-zit-detailer/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..4b7993c81d0f3a4a6b7da7cff674e8d7bb7f98ce --- /dev/null +++ b/new-fantasy-core-ill-flux-pony-sdxl-zit-detailer/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1fe5f7c36cfe6a10f9a6dd5c22362f6c18edbcf83ef7af5d8d8b9e44e4e6d993 +size 2165310 diff --git a/new-fantasy-core-ill-flux-pony-sdxl-zit-detailer/example_images/04_image.jpeg b/new-fantasy-core-ill-flux-pony-sdxl-zit-detailer/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..7e8ae5a2edd4cd1b253fa3c85b8a5fd67c41fa53 --- /dev/null +++ b/new-fantasy-core-ill-flux-pony-sdxl-zit-detailer/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab39827c0b3adf16e6f98d7ddb90eaacc77cb6a10bc81db38a27b272ff800588 +size 2108505 diff --git a/new-fantasy-core-ill-flux-pony-sdxl-zit-detailer/example_images/05_image.jpeg b/new-fantasy-core-ill-flux-pony-sdxl-zit-detailer/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..7d28fabc18a2419c5cfcef29d6440ba0bc95bf5d --- /dev/null +++ b/new-fantasy-core-ill-flux-pony-sdxl-zit-detailer/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:273c5f0bb691a07e243f43301b498fab761a0afad3dffdeaffffed68ab125fcf +size 2005930 diff --git a/new-fantasy-core-ill-flux-pony-sdxl-zit-detailer/example_images/06_image.jpeg b/new-fantasy-core-ill-flux-pony-sdxl-zit-detailer/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..f5821d685623cce4ff2f55f3d4bdfb6dbf389735 --- /dev/null +++ b/new-fantasy-core-ill-flux-pony-sdxl-zit-detailer/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9812d0663ac147b8d95a9c42e1b78d7a17270ef261c737af37b4b5071a21d92e +size 2376564 diff --git a/new-fantasy-core-ill-flux-pony-sdxl-zit-detailer/info.txt b/new-fantasy-core-ill-flux-pony-sdxl-zit-detailer/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..2c4bdcf785e1daca730bd5ebad6abb6592388ac7 --- /dev/null +++ b/new-fantasy-core-ill-flux-pony-sdxl-zit-detailer/info.txt @@ -0,0 +1,346 @@ +═══════════════════════════════════════════════════════════════════════════════ + NEW FANTASY CORE - ILL-FLUX-PONY-SDXL- ZIT(DETAILER) +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Detail / anatomy fixer + Author VisionaryAI_Studio + Source https://civitai.com/models/810000 + Provenance CivitAI · model 810000 · version 3291804 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File NewFantasyCore_Krea2.safetensors + Version NFC - Krea2 + Size 218.0 MB + Format SafeTensor / unknown precision + SHA256 eab3cd9df8b5673cc6f5895bfbf4fd3814eb0df4d3fe2b6abd1b43d2398ee374 + Published 2026-09-03 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + No trigger word. The LoRA applies as soon as it is loaded — + strength is the only way to dial the effect up or down. + +─────────────────────────────────────────────────────────────────────────────── + RECOMMENDED SETTINGS +─────────────────────────────────────────────────────────────────────────────── + The author did not publish specific settings. + No author settings were published. Community baseline: steps Turbo: 8 + (official README and HF card CLI example; the ComfyUI template ships 8). + Raw: 52 (official README). The krea-2 CLI's generic default is 28. + Community Turbo LoRAs commonly suggest 8-12 (e.g. the Identity Edit card + recommends Turbo at 8-12, or Raw at 20 for removals). A community 4-step + LoRA on Turbo makes 4 valid; its author calls 2 steps an out-of-spec + preview mode., cfg Turbo: CFG disabled - guidance_scale 0.0 in diffusers + and --cfg 0.0 in the krea-2 CLI, which is cfg 1.0 in ComfyUI's KSampler + (the official template ships cfg 1). These are the same thing; do not set + ComfyUI cfg to 0. Raw: CFG 3.5 (official README; CLI generic default 4.5). + Negative prompting on Turbo is effectively unused - the official templates + route ConditioningZeroOut into the negative slot., sampler ComfyUI: + KSampler with sampler euler, scheduler simple, denoise 1.0 (verified in + image_krea2_turbo_t2i.json: [seed, 'randomize', 8, 1, 'euler', 'simple', + 1]). Timestep shift mu = 1.15 for Turbo; ComfyUI bakes this in via + Krea2.sampling_settings = {"multiplier": 1.0, "shift": 1.15}, so the plain + t2i template needs no ModelSampling node, while the style-reference + template adds ModelSamplingFlux at max_shift 1.15 / base_shift 0.5. The CLI + exposes --mu 1.15 for Turbo; Raw uses a resolution-derived mu interpolated + between --y1 0.5 and --y2 1.15 (ai-toolkit mirrors this as base_shift 0.5 / + max_shift 1.15 over image_seq_len 256-6400, exponential time shift)., + resolution 1024x1024 default. Turbo: 1K-2K (the official CLI example + generates 2048x2048; ComfyUI's ResolutionSelector defaults to 1 megapixel, + set 2.0 for 2K). Raw: "trained to generate upto 1k resolution". + Width/height are padded up to a multiple of 16., lora_strength 1.0 for + Krea's nine official style LoRAs (docs.comfy.org lists Recommended Strength + 1.0 for every one), though the shipped t2i template's LoraLoaderModelOnly + is set to 0.8. The style-reference LoRA is used at 1.0 in its template. The + 4-step distill LoRA documents 1.0 as its trained point, 1.0-1.5 as coherent + extrapolation and above ~1.5 as speckle artefacting. Community subject + LoRAs are typically recommended around 0.8-1.2 and dropped toward 0.6-0.9 + when several are stacked.. + +─────────────────────────────────────────────────────────────────────────────── + ADOPTION & POPULARITY +─────────────────────────────────────────────────────────────────────────────── + Downloads 22,981 + Rating 100% positive (2190/2192) + Comments 0 + Published 2 days ago + Download rate 96.0/day since release + + Adoption rank #25 of 150 in this collection (High) + Momentum rank #29 of 150 by download rate (Trending · fresh release) + + Adoption tier is the percentile of total downloads within this collection; + momentum is the percentile of downloads-per-day since release. Momentum is + ranked rather than measured against a fixed rate, because this base model + is itself new and every LoRA looks fast on an absolute scale. Both numbers + are CivitAI's, read on the scrape date at the foot of this file. + +─────────────────────────────────────────────────────────────────────────────── + CHAINING / STACKING +─────────────────────────────────────────────────────────────────────────────── + Stacks with others not stated + + The author gave no chaining guidance. The rules below are the + general Krea 2 ones, not anything specific to this LoRA. + + General Krea 2 rules that apply here (see REFERENCE.txt): + + • The most important rule is not about chaining but about which + checkpoint you chain onto: the official README says in bold "TRAIN on + Raw and RUN on Turbo", and "We highly recommend using RAW for training + LoRAs and applying them on Turbo for inference." LoRAs trained on Raw + are stated to work well on Turbo. + • Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 + LoRAs are DiT-only, so there is no CLIP side to chain and the text- + encoder branch stays untouched however many LoRAs you add. + • Krea 2 LoRAs stack in practice and the community treats multi-LoRA as + routine - style + character, style + slider, realism + character. + Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one + writes "I recently stacked over 20 of my own LoRAs on a single + generation and the output still came out great." + • Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge + on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is + loaded. The CivitAI 12GB training guide also notes rank affects + composability: "rank 8 produced a lighter style influence that was + easier to stack with other LoRAs, rank 16 provided a useful balance + between concept strength and composability, and rank 32 was better + suited to a precise object or concept intended to take priority over + other stacked LoRAs." + • Accelerator LoRAs are position-sensitive in intent, not in node order: + a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a + sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to- + Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is + explicitly recommended by the 4-step LoRA's author. + • Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI + (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card + says "Keep mu = 1.15. The training targets are anchored to that grid; a + different shift moves the [sampler off them]"; the same reasoning + applies to any Turbo-trained adapter in the chain. + • Reference/edit LoRAs are designed to be stacked under ordinary subject + and style LoRAs - the Identity Edit card states "Composes with your + LoRAs: character/body/style LoRAs stack on top and steer the prior" and + recommends a stacked subject LoRA when unusual subjects drift toward + the base prior. + • Krea 2 LoRAs cannot be chained with LoRAs for any other base model. + Many CivitAI pages carry multi-model titles like "(Krea-2 + ZIT)" or + "[Flux | ZIB | Krea2]" - those are separate files for separate + architectures shipped under one model page, not a combined adapter. + • To combine two full Krea 2 checkpoints rather than adapters, ComfyUI + ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per- + section weights for first., tmlp., txtmlp., tproj., + txtfusion.projector, txtfusion.layerwise_blocks.0-1, + txtfusion.refiner_blocks.0-1, blocks.0-27 and last. + • The author never states which weight family this targets. Silent + degradation, never a hard error, between Raw and Turbo. The two share + one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" + - so every LoRA tensor shape matches on both and ComfyUI/diffuser + +─────────────────────────────────────────────────────────────────────────────── + PROS +─────────────────────────────────────────────────────────────────────────────── + + Widely adopted — 22,981 downloads. [source: CivitAI stats] + + Positively received — 2,190 thumbs-up, 2 down. [source: CivitAI stats] + + Compact at 218 MB — cheap to stack with other LoRAs. [source: file + size] + +─────────────────────────────────────────────────────────────────────────────── + CONS & CAVEATS +─────────────────────────────────────────────────────────────────────────────── + - No trigger word — the LoRA is always on once loaded, so strength is the + only control. [source: model metadata] + - Author never states which generation mode it was trained for; test + against your own workflow before relying on it. [source: description] + +─────────────────────────────────────────────────────────────────────────────── + COMPATIBILITY WARNING +─────────────────────────────────────────────────────────────────────────────── + Silent degradation, never a hard error, between Raw and Turbo. The two + share one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" - so + every LoRA tensor shape matches on both and ComfyUI/diffusers will load an + adapter against either without complaint. The lvladikov 4-step LoRA card + states this for the reverse direction: "Every layer it targets also exists + in Krea 2 Raw, so it will load there without complaint - but that is a side + effect of the shared architecture, not a supported mode", and reports + mixed, subject-dependent results with malformed subjects (duplicated heads, + fused limbs) at low step counts, needing 14+ steps with Raw's CFG on before + output is coherent. The intended direction (train on Raw, run on Turbo) is + officially blessed. The more common practical failure is a settings + mismatch rather than a weight mismatch: running a Turbo-targeted LoRA at + Raw's 52 steps + CFG 3.5, or leaving CFG enabled on Turbo, degrades output + with no warning. Loading a Krea 2 LoRA against a different model family + entirely (Flux, Qwen-Image, Z-Image) is also not an error - ComfyUI logs + "lora key not loaded: {key}" per key and generates as if no LoRA were + applied. + +─────────────────────────────────────────────────────────────────────────────── + AUTHOR'S DESCRIPTION +─────────────────────────────────────────────────────────────────────────────── + NewFantasyCore is trained on a large set of dark fantasy images to upgrade + your images with a dark style + Hey, by now New Fantasy Core has turned into a bit of chaos. Please read + the quick view so you know what to expect, since by now everyone has become + quite unique in their own ways due to the different ways of creating the + LoRA. + 📌 Quick Overview + ✦ NewFantasyCoreV2 - Krea2 [Anime Image set ] + ✦ NewFantasyCoreV2 - Z-Image [upgrade image set ] + ✦ NewFantasyCore - Z-Image [Most trained in anime and hyper-realism ] + ✦ NewFantasyCoreV4.5 - FLUX [mix between anime and semi-realistic + recommended! ] + ✦ NewFantasyCoreV4 - FLUX [Is a detailed merge of all my LoRAs, detailer + recommended! ] + ✦ NewFantasyCoreV3 - FLUX [All-rounder, beautiful dark training set, + recommended! ] + ✦ NewFantasyCoreV2 - FLUX [Strongly trained on CGI images] + ✦ NewFantasyCoreV1 - FLUX [The beginning - mix between semi-realistic and + realistic] + ✦ NewFantasyCoreV4 - ILL [Anime mix] + ✦ NewFantasyCoreV1 - ILL [FLUX V3 Image set, fantasy style, style + checkpoint-dependent] + ✦ NewFantasyCoreV1 - PONY [The beginning - mix between semi-realistic and + realistic] + ✦ NewFantasyCoreV1 - SDXL [The beginning - mix between semi-realistic and + realistic] + Features & Settings: FLUX (Update in Progress) ✔ Recommended LoRA Strength: + please see the info card for your version + ✔ Trigger: (Pay attention to the version details on the right to see if a + trigger word is required.) + ✔ Settings: CFG, STEPS, and SAMPLER are defined by the checkpoint and your + workflow + ✔ Style: different + ✔ TIP V4: if it is too messy, reduce the strength V4: errors may occur with + certain LoRA loaders in Comfyui, as I also included PONY and ILL in my merg + attempts ❗❗❗ + + Example prompts to get you started This is a highly detailed, digital + artwork featuring a + The image is a highly detailed, hyper-realistic digital artwork featuring a + The image is a digital artwork in a fantastical + This is a high-resolution photograph of a + Features & Settings: ILL + ✔ Recommended LoRA Strength: please see the info card for your version + ✔ Trigger: Newfantasycore + ✔ Settings: CFG, STEPS, and SAMPLER are defined by the checkpoint and your + workflow + ✔ Style: likes anime + Example Prompts to Get You Started + Positiv: Depends on the checkpoint, fantasy themes Negativ: Depends on the + checkpoint + Features & Settings: PONY (Update in Progress) + ✔ Recommended LoRA Strength: [0.5 – 0.7] + ✔ Trigger: (Pay attention to the version details on the right to see if a + trigger word is required.) + ✔ Settings: CFG, STEPS, and SAMPLER are defined by the checkpoint and your + workflow. + ✔ Style: Depends on the checkpoint. Semi-realistic and realistic + Example Prompts to Get You Started + Positiv: Depends on the checkpoint, fantasy themes, score_9, score_8_up, + score_7_up, score_6_up, score_5_up, Negativ: Depends on the checkpoint + Features & Settings: SDXL (Update in Progress) + ✔ Recommended LoRA Strength: [0.5 – 0.7] + ✔ Trigger: (Pay attention to the version details on the right to see if a + trigger word is required) + ✔ Settings: CFG, STEPS, and SAMPLER are defined by the checkpoint and your + workflow + ✔ Style: Depends on the checkpoint. Semi-realistic and realistic + Example Prompts to Get You Started + Positiv: Depends on the checkpoint, fantasy themes, Negativ: Depends on the + checkpoint + 📢 Feedback & Support 💬 Share your results or ask questions ✨ Suggestions + for improvements? Always welcome! 🧠 Workflow in images 👍 Like + 🔹 Enjoy generating! 🔹 + +─────────────────────────────────────────────────────────────────────────────── + EXAMPLE MEDIA +─────────────────────────────────────────────────────────────────────────────── + 6 preview(s) in ./example_images/ — community results for + this LoRA. Videos keep a matching _poster.jpg still frame. + + No prompts are recorded below: CivitAI's public API flags these posts as + having generation metadata but does not return it, so the prompts cannot be + scraped. Open the model page to read them. + + 01_image.jpeg (image, 864x1536) + 02_image.jpeg (image, 864x1536) + 03_image.jpeg (image, 864x1536) + 04_image.jpeg (image, 864x1536) + 05_image.jpeg (image, 864x1536) + 06_image.jpeg (image, 864x1536) + +─────────────────────────────────────────────────────────────────────────────── + HOW TO USE (COMFYUI) +─────────────────────────────────────────────────────────────────────────────── + 1. Update ComfyUI to at least v0.26.0 (released 2026-06-23), which added + local Krea2 support. For reference-image / style-reference / edit LoRAs you + need at least v0.29.0 (2026-07-29), which added regular and timestep-zero + reference images for Krea 2. ModelMergeKrea2 arrived in v0.27.0 + (2026-06-30). + + 2. Put the base stack in place (files from huggingface.co/Comfy-Org/Krea-2, + which is not gated): + - ComfyUI/models/diffusion_models/krea2_turbo_fp8_scaled.safetensors + (documented recommended default; the repo also holds krea2_turbo_bf16, + _int8_convrot, _mxfp8, _nvfp4 and krea2_raw_bf16 / _fp8_scaled / + _int8_convrot) + - ComfyUI/models/text_encoders/qwen3vl_4b_fp8_scaled.safetensors (bf16 + also available) + - ComfyUI/models/vae/qwen_image_vae.safetensors + + 3. Put the LoRA .safetensors in ComfyUI/models/loras/ - flat, no per-model + subfolder required. + + 4. Build or load the graph. The official image_krea2_turbo_t2i.json + template is: UNETLoader (krea2_turbo_fp8_scaled.safetensors) -> + LoraLoaderModelOnly -> KSampler; CLIPLoader + (qwen3vl_4b_fp8_scaled.safetensors, type "krea2") -> text encode -> + KSampler positive, with ConditioningZeroOut into negative; VAELoader + (qwen_image_vae.safetensors) -> VAEDecode -> SaveImage; EmptyLatentImage + 1024x1024 driven by ResolutionSelector. An int8 sibling template + (image_krea2_turbo_t2i_int8.json) also ships. + + 5. Load the LoRA with LoraLoaderModelOnly, not LoraLoader - both official + templates do. Krea 2 LoRAs patch the DiT only; note that the DiT's own + text_fusion / txtfusion blocks are a normal LoRA target (the official style + LoRAs patch them), but the Qwen3-VL text encoder itself is not. Set + strength 1.0 for Krea's official style LoRAs (the shipped template uses + 0.8) and add the LoRA's trigger word to the prompt. + + 6. Sampler settings: steps 8, cfg 1, sampler euler, scheduler simple, + denoise 1.0. Do not add a ModelSampling node for plain t2i - ComfyUI + already applies shift 1.15 from Krea2.sampling_settings. + + 7. For style reference, use the dedicated template: UNETLoader + krea2_turbo_int8_convrot.safetensors, LoraLoaderModelOnly + krea2_style_reference.safetensors at 1.0, ModelSamplingFlux (1.15, 0.5), + TextEncodeQwenImageEditPlus fed by LoadImage, + FluxKontextMultiReferenceLatentMethod set to index_timestep_zero, and + SamplerCustomAdvanced with KSamplerSelect euler + BasicScheduler simple/8 + + CFGGuider cfg 1. + + 8. For instruction editing (Krea 2 Identity Edit), stock nodes are not + enough: install the ComfyUI-Krea2Edit node pack, which ships its own + workflows, and run Turbo at 8-12 steps / cfg 1 (or Raw at 20 steps / cfg 3 + for removals) with LoRA strength 1.0. + + 9. Low-VRAM alternative: install city96's ComfyUI-GGUF and use a community + GGUF build (realrebelai/KREA-2_GGUFs, quantized from Turbo; + molbal/krea2-gguf, quantized from Raw) with the Unet Loader (GGUF) node in + place of UNETLoader. There is no official GGUF release. LoRA loading is + unchanged. + +─────────────────────────────────────────────────────────────────────────────── + Scraped by lorakit on 2026-09-05 00:12 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/new-fantasy-core-ill-flux-pony-sdxl-zit-detailer/metadata.json b/new-fantasy-core-ill-flux-pony-sdxl-zit-detailer/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..a49440af3fe724c7c1ee5529d97071918aeaf5fc --- /dev/null +++ b/new-fantasy-core-ill-flux-pony-sdxl-zit-detailer/metadata.json @@ -0,0 +1,130 @@ +{ + "name": "NEW FANTASY CORE - ILL-FLUX-PONY-SDXL- ZIT(detailer)", + "slug": "new-fantasy-core-ill-flux-pony-sdxl-zit-detailer", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 810000, + "version_id": 3291804, + "version_name": "NFC - Krea2", + "source": "civitai", + "url": "https://civitai.com/models/810000", + "repo_id": null, + "file_path": null, + "author": "VisionaryAI_Studio", + "category": { + "key": "detail", + "label": "Detail / anatomy fixer" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [], + "recommended_settings": {}, + "chaining": { + "author": { + "stacks": "not stated", + "notes": "", + "avoid": [], + "requires": [], + "order": "", + "strength": "", + "stated": false + }, + "general_rules": [ + "The most important rule is not about chaining but about which checkpoint you chain onto: the official README says in bold \"TRAIN on Raw and RUN on Turbo\", and \"We highly recommend using RAW for training LoRAs and applying them on Turbo for inference.\" LoRAs trained on Raw are stated to work well on Turbo.", + "Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 LoRAs are DiT-only, so there is no CLIP side to chain and the text-encoder branch stays untouched however many LoRAs you add.", + "Krea 2 LoRAs stack in practice and the community treats multi-LoRA as routine - style + character, style + slider, realism + character. Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one writes \"I recently stacked over 20 of my own LoRAs on a single generation and the output still came out great.\"", + "Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is loaded. The CivitAI 12GB training guide also notes rank affects composability: \"rank 8 produced a lighter style influence that was easier to stack with other LoRAs, rank 16 provided a useful balance between concept strength and composability, and rank 32 was better suited to a precise object or concept intended to take priority over other stacked LoRAs.\"", + "Accelerator LoRAs are position-sensitive in intent, not in node order: a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to-Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is explicitly recommended by the 4-step LoRA's author.", + "Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card says \"Keep mu = 1.15. The training targets are anchored to that grid; a different shift moves the [sampler off them]\"; the same reasoning applies to any Turbo-trained adapter in the chain.", + "Reference/edit LoRAs are designed to be stacked under ordinary subject and style LoRAs - the Identity Edit card states \"Composes with your LoRAs: character/body/style LoRAs stack on top and steer the prior\" and recommends a stacked subject LoRA when unusual subjects drift toward the base prior.", + "Krea 2 LoRAs cannot be chained with LoRAs for any other base model. Many CivitAI pages carry multi-model titles like \"(Krea-2 + ZIT)\" or \"[Flux | ZIB | Krea2]\" - those are separate files for separate architectures shipped under one model page, not a combined adapter.", + "To combine two full Krea 2 checkpoints rather than adapters, ComfyUI ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per-section weights for first., tmlp., txtmlp., tproj., txtfusion.projector, txtfusion.layerwise_blocks.0-1, txtfusion.refiner_blocks.0-1, blocks.0-27 and last.", + "The author never states which weight family this targets. Silent degradation, never a hard error, between Raw and Turbo. The two share one architecture - ai-toolkit's config is commented \"The reference 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)\" - so every LoRA tensor shape matches on both and ComfyUI/diffuser" + ] + }, + "pros": [ + "Widely adopted — 22,981 downloads. [source: CivitAI stats]", + "Positively received — 2,190 thumbs-up, 2 down. [source: CivitAI stats]", + "Compact at 218 MB — cheap to stack with other LoRAs. [source: file size]" + ], + "cons": [ + "No trigger word — the LoRA is always on once loaded, so strength is the only control. [source: model metadata]", + "Author never states which generation mode it was trained for; test against your own workflow before relying on it. [source: description]" + ], + "file": { + "name": "NewFantasyCore_Krea2.safetensors", + "size_bytes": 228588904, + "size_mb": 218.0, + "sha256": "eab3cd9df8b5673cc6f5895bfbf4fd3814eb0df4d3fe2b6abd1b43d2398ee374" + }, + "stats": { + "downloads": 22981, + "thumbs_up": 2190, + "thumbs_down": 2 + }, + "popularity": { + "downloads": 22981, + "thumbs_up": 2190, + "thumbs_down": 2, + "comments": 0, + "approval": "100% positive (2190/2192)", + "age_days": 1.9, + "downloads_per_day": 96.02, + "rank_by_downloads": 25, + "rank_by_momentum": 29, + "of_total": 150, + "adoption_tier": "High", + "momentum": "Trending · fresh release" + }, + "selection_score": 49261.0, + "selection_reasons": [ + "22,981 downloads", + "2,190 thumbs-up", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "864x1536", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "864x1536", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "864x1536", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "864x1536", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "864x1536", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "864x1536", + "prompt": "" + } + ], + "tags": [ + "concept", + "fantasy" + ], + "scraped_at": "2026-09-05T00:12:12+00:00", + "path": "krea-2/new-fantasy-core-ill-flux-pony-sdxl-zit-detailer" +} diff --git a/satoshi-urushihara-style/example_images/01_image.jpeg b/satoshi-urushihara-style/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..7a75525e03ec9ab28e53e771d5417e36224b147e --- /dev/null +++ b/satoshi-urushihara-style/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b35ce1775ea2ac725217ee29523e25a12d9266cdb3162551d94ff2f078e893e3 +size 4125031 diff --git a/satoshi-urushihara-style/example_images/02_image.jpeg b/satoshi-urushihara-style/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..70333275c9024d648baa623f9c9c487a547947be --- /dev/null +++ b/satoshi-urushihara-style/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25d78c6a3f6778d9bc63085f67c455a2d28c88997fd433c7fb3d77fe442d6d14 +size 3162621 diff --git a/satoshi-urushihara-style/example_images/03_image.jpeg b/satoshi-urushihara-style/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..733fc2438af24f56495e61917df559ea8fe3a705 --- /dev/null +++ b/satoshi-urushihara-style/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:37f23cd51d51a921a4f213ce80508ae52b05456d7632792cc299814a0bd99465 +size 2777060 diff --git a/satoshi-urushihara-style/example_images/04_image.jpeg b/satoshi-urushihara-style/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..ead180a189729d2f1e16250850c5cc9e270f13e7 --- /dev/null +++ b/satoshi-urushihara-style/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c7c4f7f73483fbc791a94858464bee7c735692a7d1ca17214890d62c5a0d2ca9 +size 2580755 diff --git a/satoshi-urushihara-style/example_images/05_image.jpeg b/satoshi-urushihara-style/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..52b84a7e87d53b3bbeef3779365feb901351f875 --- /dev/null +++ b/satoshi-urushihara-style/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ead71438da455bde225a454fdabff29e151a4070a586e6254a8106d4912cb88b +size 4635152 diff --git a/satoshi-urushihara-style/info.txt b/satoshi-urushihara-style/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..7007a0ee178a4bb7bd930826111cad8dd7a56428 --- /dev/null +++ b/satoshi-urushihara-style/info.txt @@ -0,0 +1,276 @@ +═══════════════════════════════════════════════════════════════════════════════ + SATOSHI URUSHIHARA STYLE +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Visual style + Author battle_bakraid + Source https://civitai.com/models/7227 + Provenance CivitAI · model 7227 · version 3080191 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File krea2_urushisato_v1_e16_r32_shrink.safetensors + Version Krea 2 + Size 197.4 MB + Format SafeTensor / unknown precision + SHA256 37d5556a095e235f68599fb84fc22b8c9b63f39638af36ad0c4b5fcfb08b4e48 + Published 2026-06-29 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + No trigger word. The LoRA applies as soon as it is loaded — + strength is the only way to dial the effect up or down. + +─────────────────────────────────────────────────────────────────────────────── + RECOMMENDED SETTINGS +─────────────────────────────────────────────────────────────────────────────── + The author did not publish specific settings. + No author settings were published. Community baseline: steps Turbo: 8 + (official README and HF card CLI example; the ComfyUI template ships 8). + Raw: 52 (official README). The krea-2 CLI's generic default is 28. + Community Turbo LoRAs commonly suggest 8-12 (e.g. the Identity Edit card + recommends Turbo at 8-12, or Raw at 20 for removals). A community 4-step + LoRA on Turbo makes 4 valid; its author calls 2 steps an out-of-spec + preview mode., cfg Turbo: CFG disabled - guidance_scale 0.0 in diffusers + and --cfg 0.0 in the krea-2 CLI, which is cfg 1.0 in ComfyUI's KSampler + (the official template ships cfg 1). These are the same thing; do not set + ComfyUI cfg to 0. Raw: CFG 3.5 (official README; CLI generic default 4.5). + Negative prompting on Turbo is effectively unused - the official templates + route ConditioningZeroOut into the negative slot., sampler ComfyUI: + KSampler with sampler euler, scheduler simple, denoise 1.0 (verified in + image_krea2_turbo_t2i.json: [seed, 'randomize', 8, 1, 'euler', 'simple', + 1]). Timestep shift mu = 1.15 for Turbo; ComfyUI bakes this in via + Krea2.sampling_settings = {"multiplier": 1.0, "shift": 1.15}, so the plain + t2i template needs no ModelSampling node, while the style-reference + template adds ModelSamplingFlux at max_shift 1.15 / base_shift 0.5. The CLI + exposes --mu 1.15 for Turbo; Raw uses a resolution-derived mu interpolated + between --y1 0.5 and --y2 1.15 (ai-toolkit mirrors this as base_shift 0.5 / + max_shift 1.15 over image_seq_len 256-6400, exponential time shift)., + resolution 1024x1024 default. Turbo: 1K-2K (the official CLI example + generates 2048x2048; ComfyUI's ResolutionSelector defaults to 1 megapixel, + set 2.0 for 2K). Raw: "trained to generate upto 1k resolution". + Width/height are padded up to a multiple of 16., lora_strength 1.0 for + Krea's nine official style LoRAs (docs.comfy.org lists Recommended Strength + 1.0 for every one), though the shipped t2i template's LoraLoaderModelOnly + is set to 0.8. The style-reference LoRA is used at 1.0 in its template. The + 4-step distill LoRA documents 1.0 as its trained point, 1.0-1.5 as coherent + extrapolation and above ~1.5 as speckle artefacting. Community subject + LoRAs are typically recommended around 0.8-1.2 and dropped toward 0.6-0.9 + when several are stacked.. + +─────────────────────────────────────────────────────────────────────────────── + ADOPTION & POPULARITY +─────────────────────────────────────────────────────────────────────────────── + Downloads 16,352 + Rating 100% positive (2387/2387) + Comments 0 + Published 67 days ago + Download rate 15.1/day since release + + Adoption rank #37 of 150 in this collection (High) + Momentum rank #117 of 150 by download rate (Long tail) + + Adoption tier is the percentile of total downloads within this collection; + momentum is the percentile of downloads-per-day since release. Momentum is + ranked rather than measured against a fixed rate, because this base model + is itself new and every LoRA looks fast on an absolute scale. Both numbers + are CivitAI's, read on the scrape date at the foot of this file. + +─────────────────────────────────────────────────────────────────────────────── + CHAINING / STACKING +─────────────────────────────────────────────────────────────────────────────── + Stacks with others not stated + + The author gave no chaining guidance. The rules below are the + general Krea 2 ones, not anything specific to this LoRA. + + General Krea 2 rules that apply here (see REFERENCE.txt): + + • The most important rule is not about chaining but about which + checkpoint you chain onto: the official README says in bold "TRAIN on + Raw and RUN on Turbo", and "We highly recommend using RAW for training + LoRAs and applying them on Turbo for inference." LoRAs trained on Raw + are stated to work well on Turbo. + • Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 + LoRAs are DiT-only, so there is no CLIP side to chain and the text- + encoder branch stays untouched however many LoRAs you add. + • Krea 2 LoRAs stack in practice and the community treats multi-LoRA as + routine - style + character, style + slider, realism + character. + Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one + writes "I recently stacked over 20 of my own LoRAs on a single + generation and the output still came out great." + • Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge + on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is + loaded. The CivitAI 12GB training guide also notes rank affects + composability: "rank 8 produced a lighter style influence that was + easier to stack with other LoRAs, rank 16 provided a useful balance + between concept strength and composability, and rank 32 was better + suited to a precise object or concept intended to take priority over + other stacked LoRAs." + • Accelerator LoRAs are position-sensitive in intent, not in node order: + a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a + sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to- + Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is + explicitly recommended by the 4-step LoRA's author. + • Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI + (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card + says "Keep mu = 1.15. The training targets are anchored to that grid; a + different shift moves the [sampler off them]"; the same reasoning + applies to any Turbo-trained adapter in the chain. + • Reference/edit LoRAs are designed to be stacked under ordinary subject + and style LoRAs - the Identity Edit card states "Composes with your + LoRAs: character/body/style LoRAs stack on top and steer the prior" and + recommends a stacked subject LoRA when unusual subjects drift toward + the base prior. + • Krea 2 LoRAs cannot be chained with LoRAs for any other base model. + Many CivitAI pages carry multi-model titles like "(Krea-2 + ZIT)" or + "[Flux | ZIB | Krea2]" - those are separate files for separate + architectures shipped under one model page, not a combined adapter. + • To combine two full Krea 2 checkpoints rather than adapters, ComfyUI + ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per- + section weights for first., tmlp., txtmlp., tproj., + txtfusion.projector, txtfusion.layerwise_blocks.0-1, + txtfusion.refiner_blocks.0-1, blocks.0-27 and last. + • The author never states which weight family this targets. Silent + degradation, never a hard error, between Raw and Turbo. The two share + one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" + - so every LoRA tensor shape matches on both and ComfyUI/diffuser + +─────────────────────────────────────────────────────────────────────────────── + PROS +─────────────────────────────────────────────────────────────────────────────── + + Widely adopted — 16,352 downloads. [source: CivitAI stats] + + Positively received — 2,387 thumbs-up with no down-votes. [source: + CivitAI stats] + + Compact at 197 MB — cheap to stack with other LoRAs. [source: file + size] + +─────────────────────────────────────────────────────────────────────────────── + CONS & CAVEATS +─────────────────────────────────────────────────────────────────────────────── + - No trigger word — the LoRA is always on once loaded, so strength is the + only control. [source: model metadata] + - Author never states which generation mode it was trained for; test + against your own workflow before relying on it. [source: description] + - Thin documentation — expect trial and error. [source: description + length] + +─────────────────────────────────────────────────────────────────────────────── + COMPATIBILITY WARNING +─────────────────────────────────────────────────────────────────────────────── + Silent degradation, never a hard error, between Raw and Turbo. The two + share one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" - so + every LoRA tensor shape matches on both and ComfyUI/diffusers will load an + adapter against either without complaint. The lvladikov 4-step LoRA card + states this for the reverse direction: "Every layer it targets also exists + in Krea 2 Raw, so it will load there without complaint - but that is a side + effect of the shared architecture, not a supported mode", and reports + mixed, subject-dependent results with malformed subjects (duplicated heads, + fused limbs) at low step counts, needing 14+ steps with Raw's CFG on before + output is coherent. The intended direction (train on Raw, run on Turbo) is + officially blessed. The more common practical failure is a settings + mismatch rather than a weight mismatch: running a Turbo-targeted LoRA at + Raw's 52 steps + CFG 3.5, or leaving CFG enabled on Turbo, degrades output + with no warning. Loading a Krea 2 LoRA against a different model family + entirely (Flux, Qwen-Image, Z-Image) is also not an error - ComfyUI logs + "lora key not loaded: {key}" per key and generates as if no LoRA were + applied. + +─────────────────────────────────────────────────────────────────────────────── + AUTHOR'S DESCRIPTION +─────────────────────────────────────────────────────────────────────────────── + for that late 80's early 90's anime aesthetic + Trained on Satoshi Urushihara artwork and OVA stills. + +─────────────────────────────────────────────────────────────────────────────── + EXAMPLE MEDIA +─────────────────────────────────────────────────────────────────────────────── + 5 preview(s) in ./example_images/ — community results for + this LoRA. Videos keep a matching _poster.jpg still frame. + + No prompts are recorded below: CivitAI's public API flags these posts as + having generation metadata but does not return it, so the prompts cannot be + scraped. Open the model page to read them. + + 01_image.jpeg (image, 1400x2120) + 02_image.jpeg (image, 1152x1728) + 03_image.jpeg (image, 1152x1728) + 04_image.jpeg (image, 1152x1728) + 05_image.jpeg (image, 1400x2120) + +─────────────────────────────────────────────────────────────────────────────── + HOW TO USE (COMFYUI) +─────────────────────────────────────────────────────────────────────────────── + 1. Update ComfyUI to at least v0.26.0 (released 2026-06-23), which added + local Krea2 support. For reference-image / style-reference / edit LoRAs you + need at least v0.29.0 (2026-07-29), which added regular and timestep-zero + reference images for Krea 2. ModelMergeKrea2 arrived in v0.27.0 + (2026-06-30). + + 2. Put the base stack in place (files from huggingface.co/Comfy-Org/Krea-2, + which is not gated): + - ComfyUI/models/diffusion_models/krea2_turbo_fp8_scaled.safetensors + (documented recommended default; the repo also holds krea2_turbo_bf16, + _int8_convrot, _mxfp8, _nvfp4 and krea2_raw_bf16 / _fp8_scaled / + _int8_convrot) + - ComfyUI/models/text_encoders/qwen3vl_4b_fp8_scaled.safetensors (bf16 + also available) + - ComfyUI/models/vae/qwen_image_vae.safetensors + + 3. Put the LoRA .safetensors in ComfyUI/models/loras/ - flat, no per-model + subfolder required. + + 4. Build or load the graph. The official image_krea2_turbo_t2i.json + template is: UNETLoader (krea2_turbo_fp8_scaled.safetensors) -> + LoraLoaderModelOnly -> KSampler; CLIPLoader + (qwen3vl_4b_fp8_scaled.safetensors, type "krea2") -> text encode -> + KSampler positive, with ConditioningZeroOut into negative; VAELoader + (qwen_image_vae.safetensors) -> VAEDecode -> SaveImage; EmptyLatentImage + 1024x1024 driven by ResolutionSelector. An int8 sibling template + (image_krea2_turbo_t2i_int8.json) also ships. + + 5. Load the LoRA with LoraLoaderModelOnly, not LoraLoader - both official + templates do. Krea 2 LoRAs patch the DiT only; note that the DiT's own + text_fusion / txtfusion blocks are a normal LoRA target (the official style + LoRAs patch them), but the Qwen3-VL text encoder itself is not. Set + strength 1.0 for Krea's official style LoRAs (the shipped template uses + 0.8) and add the LoRA's trigger word to the prompt. + + 6. Sampler settings: steps 8, cfg 1, sampler euler, scheduler simple, + denoise 1.0. Do not add a ModelSampling node for plain t2i - ComfyUI + already applies shift 1.15 from Krea2.sampling_settings. + + 7. For style reference, use the dedicated template: UNETLoader + krea2_turbo_int8_convrot.safetensors, LoraLoaderModelOnly + krea2_style_reference.safetensors at 1.0, ModelSamplingFlux (1.15, 0.5), + TextEncodeQwenImageEditPlus fed by LoadImage, + FluxKontextMultiReferenceLatentMethod set to index_timestep_zero, and + SamplerCustomAdvanced with KSamplerSelect euler + BasicScheduler simple/8 + + CFGGuider cfg 1. + + 8. For instruction editing (Krea 2 Identity Edit), stock nodes are not + enough: install the ComfyUI-Krea2Edit node pack, which ships its own + workflows, and run Turbo at 8-12 steps / cfg 1 (or Raw at 20 steps / cfg 3 + for removals) with LoRA strength 1.0. + + 9. Low-VRAM alternative: install city96's ComfyUI-GGUF and use a community + GGUF build (realrebelai/KREA-2_GGUFs, quantized from Turbo; + molbal/krea2-gguf, quantized from Raw) with the Unet Loader (GGUF) node in + place of UNETLoader. There is no official GGUF release. LoRA loading is + unchanged. + +─────────────────────────────────────────────────────────────────────────────── + Scraped by lorakit on 2026-09-05 00:12 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/satoshi-urushihara-style/krea2_urushisato_v1_e16_r32_shrink.safetensors b/satoshi-urushihara-style/krea2_urushisato_v1_e16_r32_shrink.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..3a01006cb861934286aec637efdf5d7504c75636 --- /dev/null +++ b/satoshi-urushihara-style/krea2_urushisato_v1_e16_r32_shrink.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:37d5556a095e235f68599fb84fc22b8c9b63f39638af36ad0c4b5fcfb08b4e48 +size 207001082 diff --git a/satoshi-urushihara-style/metadata.json b/satoshi-urushihara-style/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..168c1a78436635b486a17a8a8f7d6d5220dd99e9 --- /dev/null +++ b/satoshi-urushihara-style/metadata.json @@ -0,0 +1,129 @@ +{ + "name": "Satoshi Urushihara style", + "slug": "satoshi-urushihara-style", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 7227, + "version_id": 3080191, + "version_name": "Krea 2", + "source": "civitai", + "url": "https://civitai.com/models/7227", + "repo_id": null, + "file_path": null, + "author": "battle_bakraid", + "category": { + "key": "style", + "label": "Visual style" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [], + "recommended_settings": {}, + "chaining": { + "author": { + "stacks": "not stated", + "notes": "", + "avoid": [], + "requires": [], + "order": "", + "strength": "", + "stated": false + }, + "general_rules": [ + "The most important rule is not about chaining but about which checkpoint you chain onto: the official README says in bold \"TRAIN on Raw and RUN on Turbo\", and \"We highly recommend using RAW for training LoRAs and applying them on Turbo for inference.\" LoRAs trained on Raw are stated to work well on Turbo.", + "Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 LoRAs are DiT-only, so there is no CLIP side to chain and the text-encoder branch stays untouched however many LoRAs you add.", + "Krea 2 LoRAs stack in practice and the community treats multi-LoRA as routine - style + character, style + slider, realism + character. Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one writes \"I recently stacked over 20 of my own LoRAs on a single generation and the output still came out great.\"", + "Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is loaded. The CivitAI 12GB training guide also notes rank affects composability: \"rank 8 produced a lighter style influence that was easier to stack with other LoRAs, rank 16 provided a useful balance between concept strength and composability, and rank 32 was better suited to a precise object or concept intended to take priority over other stacked LoRAs.\"", + "Accelerator LoRAs are position-sensitive in intent, not in node order: a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to-Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is explicitly recommended by the 4-step LoRA's author.", + "Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card says \"Keep mu = 1.15. The training targets are anchored to that grid; a different shift moves the [sampler off them]\"; the same reasoning applies to any Turbo-trained adapter in the chain.", + "Reference/edit LoRAs are designed to be stacked under ordinary subject and style LoRAs - the Identity Edit card states \"Composes with your LoRAs: character/body/style LoRAs stack on top and steer the prior\" and recommends a stacked subject LoRA when unusual subjects drift toward the base prior.", + "Krea 2 LoRAs cannot be chained with LoRAs for any other base model. Many CivitAI pages carry multi-model titles like \"(Krea-2 + ZIT)\" or \"[Flux | ZIB | Krea2]\" - those are separate files for separate architectures shipped under one model page, not a combined adapter.", + "To combine two full Krea 2 checkpoints rather than adapters, ComfyUI ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per-section weights for first., tmlp., txtmlp., tproj., txtfusion.projector, txtfusion.layerwise_blocks.0-1, txtfusion.refiner_blocks.0-1, blocks.0-27 and last.", + "The author never states which weight family this targets. Silent degradation, never a hard error, between Raw and Turbo. The two share one architecture - ai-toolkit's config is commented \"The reference 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)\" - so every LoRA tensor shape matches on both and ComfyUI/diffuser" + ] + }, + "pros": [ + "Widely adopted — 16,352 downloads. [source: CivitAI stats]", + "Positively received — 2,387 thumbs-up with no down-votes. [source: CivitAI stats]", + "Compact at 197 MB — cheap to stack with other LoRAs. [source: file size]" + ], + "cons": [ + "No trigger word — the LoRA is always on once loaded, so strength is the only control. [source: model metadata]", + "Author never states which generation mode it was trained for; test against your own workflow before relying on it. [source: description]", + "Thin documentation — expect trial and error. [source: description length]" + ], + "file": { + "name": "krea2_urushisato_v1_e16_r32_shrink.safetensors", + "size_bytes": 207001082, + "size_mb": 197.41, + "sha256": "37d5556a095e235f68599fb84fc22b8c9b63f39638af36ad0c4b5fcfb08b4e48" + }, + "stats": { + "downloads": 16352, + "thumbs_up": 2387, + "thumbs_down": 0 + }, + "popularity": { + "downloads": 16352, + "thumbs_up": 2387, + "thumbs_down": 0, + "comments": 0, + "approval": "100% positive (2387/2387)", + "age_days": 67.4, + "downloads_per_day": 15.08, + "rank_by_downloads": 37, + "rank_by_momentum": 117, + "of_total": 150, + "adoption_tier": "High", + "momentum": "Long tail" + }, + "selection_score": 44996.0, + "selection_reasons": [ + "16,352 downloads", + "2,387 thumbs-up", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "1400x2120", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "1152x1728", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "1152x1728", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "1152x1728", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "1400x2120", + "prompt": "" + } + ], + "tags": [ + "style", + "vintage", + "anime", + "aesthetic", + "1980s", + "1990s" + ], + "scraped_at": "2026-09-05T00:12:17+00:00", + "path": "krea-2/satoshi-urushihara-style" +} diff --git a/slavekini-aka-slave-leia-outfit-clothing-krea2-anima-il/Slavekini_Krea2_epoch_7.safetensors b/slavekini-aka-slave-leia-outfit-clothing-krea2-anima-il/Slavekini_Krea2_epoch_7.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..6ac5a7127aaa7a3e35e214b5201421458417c85b --- /dev/null +++ b/slavekini-aka-slave-leia-outfit-clothing-krea2-anima-il/Slavekini_Krea2_epoch_7.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ff593bf2f0b67a8101e3adb894ec94b5d16dfb3b7dd69fbec9df488ca772d33e +size 228588904 diff --git a/slavekini-aka-slave-leia-outfit-clothing-krea2-anima-il/example_images/01_image.jpeg b/slavekini-aka-slave-leia-outfit-clothing-krea2-anima-il/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..3cb75a82fe193e0f7a42fec1cc7d739d78291378 --- /dev/null +++ b/slavekini-aka-slave-leia-outfit-clothing-krea2-anima-il/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:993d1ad985343096fd632170cd8093070c338c4697298504509eff3cbac683d7 +size 3545700 diff --git a/slavekini-aka-slave-leia-outfit-clothing-krea2-anima-il/example_images/02_image.jpeg b/slavekini-aka-slave-leia-outfit-clothing-krea2-anima-il/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..33b29e669eb14f7579854ff3cb1aa77c6b1f8f57 --- /dev/null +++ b/slavekini-aka-slave-leia-outfit-clothing-krea2-anima-il/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dfaf8af90f1747e10f5dfe97ccddfb3e73479fe17e47d9b465b95dfbe95ad388 +size 1721096 diff --git a/slavekini-aka-slave-leia-outfit-clothing-krea2-anima-il/example_images/03_image.jpeg b/slavekini-aka-slave-leia-outfit-clothing-krea2-anima-il/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..58aba97c1e35578df9f81873a25ace8fdb50ba41 --- /dev/null +++ b/slavekini-aka-slave-leia-outfit-clothing-krea2-anima-il/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d318607bd72e9555492dd69071c01ba3074c1de97de50487f906bfda92a89775 +size 2648946 diff --git a/slavekini-aka-slave-leia-outfit-clothing-krea2-anima-il/example_images/04_image.jpeg b/slavekini-aka-slave-leia-outfit-clothing-krea2-anima-il/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..b5a503131a224e07355996115de5594ebdbd0add --- /dev/null +++ b/slavekini-aka-slave-leia-outfit-clothing-krea2-anima-il/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a042c8c93c9349d283a6ba3d175b22a5fcd20a209fe90ce899fbbdbbee73469 +size 2662898 diff --git a/slavekini-aka-slave-leia-outfit-clothing-krea2-anima-il/example_images/05_image.jpeg b/slavekini-aka-slave-leia-outfit-clothing-krea2-anima-il/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..aed0ebfaae0ee6b5d1ac8c8e704f3c263de65811 --- /dev/null +++ b/slavekini-aka-slave-leia-outfit-clothing-krea2-anima-il/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3598832028684e3abced42072f6fff0c5c80f9e2ae32ab64c3104ccadc50654a +size 2969188 diff --git a/slavekini-aka-slave-leia-outfit-clothing-krea2-anima-il/example_images/06_image.jpeg b/slavekini-aka-slave-leia-outfit-clothing-krea2-anima-il/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..ad785d951e073a14f57b588d3c445a0d18bdba6e --- /dev/null +++ b/slavekini-aka-slave-leia-outfit-clothing-krea2-anima-il/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3d9f1c5e6c52475ee0946906476d8315a1fd1f5310ea7fe01053acb8303bfc0c +size 2315347 diff --git a/slavekini-aka-slave-leia-outfit-clothing-krea2-anima-il/info.txt b/slavekini-aka-slave-leia-outfit-clothing-krea2-anima-il/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..47579670c463c3bd5469d86881403ed0ec7b0c55 --- /dev/null +++ b/slavekini-aka-slave-leia-outfit-clothing-krea2-anima-il/info.txt @@ -0,0 +1,376 @@ +═══════════════════════════════════════════════════════════════════════════════ + SLAVEKINI, AKA SLAVE LEIA OUTFIT - CLOTHING [KREA2, ANIMA, IL, PONY, SDXL] +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Photorealism + Author cranpeach + Source https://civitai.com/models/7913 + Provenance CivitAI · model 7913 · version 3151337 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File Slavekini_Krea2_epoch_7.safetensors + Version Krea2 + Size 218.0 MB + Format SafeTensor / unknown precision + SHA256 ff593bf2f0b67a8101e3adb894ec94b5d16dfb3b7dd69fbec9df488ca772d33e + Published 2026-07-21 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + › "wearing a green slavekini top with burgundy loin cloth" + › "gold choker, gold armlet, and a single wrist cuff" + › "hair up in a gold hair ornament, with hair in braid over shoulder with + gold hair tie" + +─────────────────────────────────────────────────────────────────────────────── + RECOMMENDED SETTINGS +─────────────────────────────────────────────────────────────────────────────── + Taken verbatim from the author's description: + + LoRA strength 0.8–1.0 + +─────────────────────────────────────────────────────────────────────────────── + ADOPTION & POPULARITY +─────────────────────────────────────────────────────────────────────────────── + Downloads 29,303 + Rating 100% positive (3234/3235) + Comments 0 + Published 45 days ago + Download rate 26.4/day since release + + Adoption rank #20 of 150 in this collection (High) + Momentum rank #84 of 150 by download rate (Long tail) + + Adoption tier is the percentile of total downloads within this collection; + momentum is the percentile of downloads-per-day since release. Momentum is + ranked rather than measured against a fixed rate, because this base model + is itself new and every LoRA looks fast on an absolute scale. Both numbers + are CivitAI's, read on the scrape date at the foot of this file. + +─────────────────────────────────────────────────────────────────────────────── + CHAINING / STACKING +─────────────────────────────────────────────────────────────────────────────── + Stacks with others not stated + + The author gave no chaining guidance. The rules below are the + general Krea 2 ones, not anything specific to this LoRA. + + General Krea 2 rules that apply here (see REFERENCE.txt): + + • The most important rule is not about chaining but about which + checkpoint you chain onto: the official README says in bold "TRAIN on + Raw and RUN on Turbo", and "We highly recommend using RAW for training + LoRAs and applying them on Turbo for inference." LoRAs trained on Raw + are stated to work well on Turbo. + • Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 + LoRAs are DiT-only, so there is no CLIP side to chain and the text- + encoder branch stays untouched however many LoRAs you add. + • Krea 2 LoRAs stack in practice and the community treats multi-LoRA as + routine - style + character, style + slider, realism + character. + Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one + writes "I recently stacked over 20 of my own LoRAs on a single + generation and the output still came out great." + • Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge + on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is + loaded. The CivitAI 12GB training guide also notes rank affects + composability: "rank 8 produced a lighter style influence that was + easier to stack with other LoRAs, rank 16 provided a useful balance + between concept strength and composability, and rank 32 was better + suited to a precise object or concept intended to take priority over + other stacked LoRAs." + • Accelerator LoRAs are position-sensitive in intent, not in node order: + a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a + sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to- + Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is + explicitly recommended by the 4-step LoRA's author. + • Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI + (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card + says "Keep mu = 1.15. The training targets are anchored to that grid; a + different shift moves the [sampler off them]"; the same reasoning + applies to any Turbo-trained adapter in the chain. + • Reference/edit LoRAs are designed to be stacked under ordinary subject + and style LoRAs - the Identity Edit card states "Composes with your + LoRAs: character/body/style LoRAs stack on top and steer the prior" and + recommends a stacked subject LoRA when unusual subjects drift toward + the base prior. + • Krea 2 LoRAs cannot be chained with LoRAs for any other base model. + Many CivitAI pages carry multi-model titles like "(Krea-2 + ZIT)" or + "[Flux | ZIB | Krea2]" - those are separate files for separate + architectures shipped under one model page, not a combined adapter. + • To combine two full Krea 2 checkpoints rather than adapters, ComfyUI + ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per- + section weights for first., tmlp., txtmlp., tproj., + txtfusion.projector, txtfusion.layerwise_blocks.0-1, + txtfusion.refiner_blocks.0-1, blocks.0-27 and last. + • The author never states which weight family this targets. Silent + degradation, never a hard error, between Raw and Turbo. The two share + one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" + - so every LoRA tensor shape matches on both and ComfyUI/diffuser + +─────────────────────────────────────────────────────────────────────────────── + PROS +─────────────────────────────────────────────────────────────────────────────── + + Widely adopted — 29,303 downloads. [source: CivitAI stats] + + Positively received — 3,234 thumbs-up, 1 down. [source: CivitAI stats] + + Explicit trigger word(s) — wearing a green slavekini top with burgundy + loin cloth, gold choker, gold armlet, and a single wrist cuff, hair up + in a gold hair ornament, with hair in braid over shoulder with gold + hair tie — so the effect can be turned on and off from the prompt. + [source: model metadata] + + Compact at 218 MB — cheap to stack with other LoRAs. [source: file + size] + + Author published concrete recommended settings. [source: description] + +─────────────────────────────────────────────────────────────────────────────── + CONS & CAVEATS +─────────────────────────────────────────────────────────────────────────────── + - Needs a reduced strength; the default 1.0 is likely too strong. + [source: description] + - Author references NSFW output — review before using in shared work. + [source: description] + - Author never states which generation mode it was trained for; test + against your own workflow before relying on it. [source: description] + +─────────────────────────────────────────────────────────────────────────────── + COMPATIBILITY WARNING +─────────────────────────────────────────────────────────────────────────────── + Silent degradation, never a hard error, between Raw and Turbo. The two + share one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" - so + every LoRA tensor shape matches on both and ComfyUI/diffusers will load an + adapter against either without complaint. The lvladikov 4-step LoRA card + states this for the reverse direction: "Every layer it targets also exists + in Krea 2 Raw, so it will load there without complaint - but that is a side + effect of the shared architecture, not a supported mode", and reports + mixed, subject-dependent results with malformed subjects (duplicated heads, + fused limbs) at low step counts, needing 14+ steps with Raw's CFG on before + output is coherent. The intended direction (train on Raw, run on Turbo) is + officially blessed. The more common practical failure is a settings + mismatch rather than a weight mismatch: running a Turbo-targeted LoRA at + Raw's 52 steps + CFG 3.5, or leaving CFG enabled on Turbo, degrades output + with no warning. Loading a Krea 2 LoRA against a different model family + entirely (Flux, Qwen-Image, Z-Image) is also not an error - ComfyUI logs + "lora key not loaded: {key}" per key and generates as if no LoRA were + applied. + +─────────────────────────────────────────────────────────────────────────────── + AUTHOR'S DESCRIPTION +─────────────────────────────────────────────────────────────────────────────── + Krea 2 Update 7/21/2026 + Tag usage: wearing a slavekini , with trim , loin + cloth , gold choker , gold armlet , a single wrist cuff , hair up in a gold + hair ornament , with hair in braid over shoulder with gold hair tie + Suggested strength: .8-1.0 + Using slave leia outfit does about the same as slavekini but does carry a + bit more strength to it - just beware using the leia tag runs the risk of + giving your character Princess Leia hair buns. May still have to prompt for + the braid and choker. Suggest using shiny slavekini top if you like + shiny stuff as it is in the dataset. + As usual use lower strength for more flexibility/use with multiple loras, + use high strength for more accuracy + Anima Base Update 6/3/2026 + Suggested tag: slave leia outfit for full outfit + Seperate outfit parts: slavekini , pelvic curtain , collar , armlet , hair + ornament , braid - can also prompt for slavekini + Entirely new dataset trained on Anima Base 1.0. + Works surprisingly similar to previous SDXL-based models, trained on the + tag slavekini , but if you want all the bells and whistles of the outfit, + you can use the tag slave leia outfit . + The tag slavekini is likely to create something new based off the + characters outfit. + Illustrious Update 12/31/24 + Suggested tag: slave leia outfit for full outfit + Seperate outfit parts: slavekini , pelvic curtain , collar , armlet , hair + ornament , braid - can also prompt for slavekini + Same notes for Pony model applies here, trained on the tag slavekini , but + if you want all the bells and whistles of the outfit, I suggest just using + slave leia outfit . Works on Illustrious as well as NoobAI. + The tag slavekini works similar to the SDXL Update, where it's more likely + to create something new based off the characters outfit, and will not make + their hair in a ponytail. It also requires you to prompt for pelvic curtain + if wanted. + Pony XL Update 4/1/24 + Trained on Pony XL base model. Suggested tag: slave leia outfit + While I did train this LoRa on the tag slavekini , if you want all the + bells and whistles of the outfit, I suggest just using slave leia outfit . + The Pony model is actually already trained on this outfit, there's all the + right tokens floating around, this Lora will just help make it more spot + on. + The tag slavekini works similar to the SDXL Update, where it's more likely + to create something new based off the characters outfit, and will not make + their hair in a ponytail. It also requires you to prompt for pelvic curtain + if wanted. + + Update: You can now generate and upload images using the slave leia tag + again. Shoutouts to the swift work of the Civitai support discord! + SDXL Update: 10/5/23 + What you need to know: + Prompt: wearing a slavekini - More flexible to the character you are + putting it on, will change colors more willingly (i.e red slavekini ) and + their hair shouldn't have the hair ornament . May have to add pelvic + curtain + Prompt: wearing a slave leia outfit - Natural token built into SDXL that + has also been trained into this lora - very willing to reproduce the outfit + including the braid, hair ornament, makeup, pelvic curtain, armlet, + bracelet, brown hair, cleavage without prompting for it. These can also be + prompted away in the negative prompt, albeit a bit harder to get rid of. + Suggested weight: 1 + Tags you can prompt for in positive or negative prompt: + pelvic curtain, arm up, jewelry, collar, chain, holding chain, leash, + braid, armlet, bracelet, from behind, long hair, cleavage, small breasts, + medium breasts, large breasts, huge breasts, no makeup, makeup, eyeliner, + eyeshadow, eyelashes, lipstick, navel piercing, earrings, kneeling, + realistic This model is less willing to go topless but it should work if + you remove the prompt: slavekini or slave leia outfit . bottomless seems + impossible despite having been trained on it, good luck. + Backstory: This is technically a merge of two LoCon's, one trained on the + token slavekini and the second on it's natural token in SDXL slave leia + outfit . I initially started this with the intention of no longer using + slavekini when I found there is a token already trained into SDXL for slave + leia outfit . But what I found is that the natural token is far too willing + to add hair ornament and braid as well as Princess leias hair buns, crown, + or goggles, likely just from the leia token alone. I went in and tagged + everything that had the hair ornament , tried inpainting the hair ornament + out of several training images, tried inpainting women without makeup to + get it less focused on the makeup, added prompts for different kinds of + makeup they were wearing, but to no avail, it kept coming back and wouldn't + even prompt away in the negative prompt with extra weights added. While + some may want this, I prefer the flexibility of my previous model where the + style of the top would change depending on the character it was on. That is + why there are now two available prompts to activate this lora. + I've found that you can also use slave leia bikini and slave leia cosplay + Examples: + + Update: 2/14/23 - 9MB file size! (shoutouts to LuisaP!) + + - Removed chest tattoo from dataset that kept showing up in gens. + + - Mild changes to the dataset- hopefully able to do upper body only as well + as full body. + + - Armlet location should more often appear only on the right side. + + - Might be a pinch less flexible than the previous version based on + testing. + + Lots of trial and error but finally getting the hang of LORAs. + First ever public release so be gentle! + Prompt: slavekini should be all you need to get the essentials. + This thing can do nsfw, different poses, angles, topless, bottomless, all + that good stuff. Prompting "from behind" likes to make yoga pants so may + have to prompt for pelvic curtain. Can also negative prompt to get rid of + certain things. If using with other character loras, honestly a lot depends + on the flexibility of their model. Trained on 1.5 clip 1 but works great + for anime or realistic models and clip 2. Tested on AOM2, Deliberate, among + others. + Optional prompts that do fun stuff: chain, bound, dress removed, + bottomless, topless, leash collar, pelvic curtain + +─────────────────────────────────────────────────────────────────────────────── + VERSION NOTES +─────────────────────────────────────────────────────────────────────────────── + Tag usage: wearing a slavekini , with trim , loin + cloth , gold choker , gold armlet , a single wrist cuff , hair up in a gold + hair ornament , with hair in braid over shoulder with gold hair tie + Suggested strength: 0.8-1.0 + Using slave leia outfit does about the same as slavekini but does carry a + bit more strength to it - just beware using the leia tag runs the risk of + giving your character Princess Leia hair buns. May still have to prompt for + the braid and choker. Suggest using shiny slavekini top if you like + shiny stuff as it is in the dataset. + As usual use lower strength for more flexibility/use with multiple loras, + use high strength for more accuracy + +─────────────────────────────────────────────────────────────────────────────── + EXAMPLE MEDIA +─────────────────────────────────────────────────────────────────────────────── + 6 preview(s) in ./example_images/ — community results for + this LoRA. Videos keep a matching _poster.jpg still frame. + + No prompts are recorded below: CivitAI's public API flags these posts as + having generation metadata but does not return it, so the prompts cannot be + scraped. Open the model page to read them. + + 01_image.jpeg (image, 1152x1792) + 02_image.jpeg (image, 1024x1536) + 03_image.jpeg (image, 1120x1664) + 04_image.jpeg (image, 1152x1792) + 05_image.jpeg (image, 1152x1792) + 06_image.jpeg (image, 1024x1536) + +─────────────────────────────────────────────────────────────────────────────── + HOW TO USE (COMFYUI) +─────────────────────────────────────────────────────────────────────────────── + 1. Update ComfyUI to at least v0.26.0 (released 2026-06-23), which added + local Krea2 support. For reference-image / style-reference / edit LoRAs you + need at least v0.29.0 (2026-07-29), which added regular and timestep-zero + reference images for Krea 2. ModelMergeKrea2 arrived in v0.27.0 + (2026-06-30). + + 2. Put the base stack in place (files from huggingface.co/Comfy-Org/Krea-2, + which is not gated): + - ComfyUI/models/diffusion_models/krea2_turbo_fp8_scaled.safetensors + (documented recommended default; the repo also holds krea2_turbo_bf16, + _int8_convrot, _mxfp8, _nvfp4 and krea2_raw_bf16 / _fp8_scaled / + _int8_convrot) + - ComfyUI/models/text_encoders/qwen3vl_4b_fp8_scaled.safetensors (bf16 + also available) + - ComfyUI/models/vae/qwen_image_vae.safetensors + + 3. Put the LoRA .safetensors in ComfyUI/models/loras/ - flat, no per-model + subfolder required. + + 4. Build or load the graph. The official image_krea2_turbo_t2i.json + template is: UNETLoader (krea2_turbo_fp8_scaled.safetensors) -> + LoraLoaderModelOnly -> KSampler; CLIPLoader + (qwen3vl_4b_fp8_scaled.safetensors, type "krea2") -> text encode -> + KSampler positive, with ConditioningZeroOut into negative; VAELoader + (qwen_image_vae.safetensors) -> VAEDecode -> SaveImage; EmptyLatentImage + 1024x1024 driven by ResolutionSelector. An int8 sibling template + (image_krea2_turbo_t2i_int8.json) also ships. + + 5. Load the LoRA with LoraLoaderModelOnly, not LoraLoader - both official + templates do. Krea 2 LoRAs patch the DiT only; note that the DiT's own + text_fusion / txtfusion blocks are a normal LoRA target (the official style + LoRAs patch them), but the Qwen3-VL text encoder itself is not. Set + strength 1.0 for Krea's official style LoRAs (the shipped template uses + 0.8) and add the LoRA's trigger word to the prompt. + + 6. Sampler settings: steps 8, cfg 1, sampler euler, scheduler simple, + denoise 1.0. Do not add a ModelSampling node for plain t2i - ComfyUI + already applies shift 1.15 from Krea2.sampling_settings. + + 7. For style reference, use the dedicated template: UNETLoader + krea2_turbo_int8_convrot.safetensors, LoraLoaderModelOnly + krea2_style_reference.safetensors at 1.0, ModelSamplingFlux (1.15, 0.5), + TextEncodeQwenImageEditPlus fed by LoadImage, + FluxKontextMultiReferenceLatentMethod set to index_timestep_zero, and + SamplerCustomAdvanced with KSamplerSelect euler + BasicScheduler simple/8 + + CFGGuider cfg 1. + + 8. For instruction editing (Krea 2 Identity Edit), stock nodes are not + enough: install the ComfyUI-Krea2Edit node pack, which ships its own + workflows, and run Turbo at 8-12 steps / cfg 1 (or Raw at 20 steps / cfg 3 + for removals) with LoRA strength 1.0. + + 9. Low-VRAM alternative: install city96's ComfyUI-GGUF and use a community + GGUF build (realrebelai/KREA-2_GGUFs, quantized from Turbo; + molbal/krea2-gguf, quantized from Raw) with the Unet Loader (GGUF) node in + place of UNETLoader. There is no official GGUF release. LoRA loading is + unchanged. + +─────────────────────────────────────────────────────────────────────────────── + Scraped by lorakit on 2026-09-05 00:11 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/slavekini-aka-slave-leia-outfit-clothing-krea2-anima-il/metadata.json b/slavekini-aka-slave-leia-outfit-clothing-krea2-anima-il/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..ab600fff04384b166fa8838c62a53437e3429945 --- /dev/null +++ b/slavekini-aka-slave-leia-outfit-clothing-krea2-anima-il/metadata.json @@ -0,0 +1,147 @@ +{ + "name": "Slavekini, aka Slave Leia Outfit - Clothing [Krea2, Anima, IL, Pony, SDXL]", + "slug": "slavekini-aka-slave-leia-outfit-clothing-krea2-anima-il", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 7913, + "version_id": 3151337, + "version_name": "Krea2", + "source": "civitai", + "url": "https://civitai.com/models/7913", + "repo_id": null, + "file_path": null, + "author": "cranpeach", + "category": { + "key": "photoreal", + "label": "Photorealism" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [ + "wearing a green slavekini top with burgundy loin cloth", + "gold choker, gold armlet, and a single wrist cuff", + "hair up in a gold hair ornament, with hair in braid over shoulder with gold hair tie" + ], + "recommended_settings": { + "LoRA strength": "0.8–1.0" + }, + "chaining": { + "author": { + "stacks": "not stated", + "notes": "", + "avoid": [], + "requires": [], + "order": "", + "strength": "", + "stated": false + }, + "general_rules": [ + "The most important rule is not about chaining but about which checkpoint you chain onto: the official README says in bold \"TRAIN on Raw and RUN on Turbo\", and \"We highly recommend using RAW for training LoRAs and applying them on Turbo for inference.\" LoRAs trained on Raw are stated to work well on Turbo.", + "Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 LoRAs are DiT-only, so there is no CLIP side to chain and the text-encoder branch stays untouched however many LoRAs you add.", + "Krea 2 LoRAs stack in practice and the community treats multi-LoRA as routine - style + character, style + slider, realism + character. Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one writes \"I recently stacked over 20 of my own LoRAs on a single generation and the output still came out great.\"", + "Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is loaded. The CivitAI 12GB training guide also notes rank affects composability: \"rank 8 produced a lighter style influence that was easier to stack with other LoRAs, rank 16 provided a useful balance between concept strength and composability, and rank 32 was better suited to a precise object or concept intended to take priority over other stacked LoRAs.\"", + "Accelerator LoRAs are position-sensitive in intent, not in node order: a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to-Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is explicitly recommended by the 4-step LoRA's author.", + "Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card says \"Keep mu = 1.15. The training targets are anchored to that grid; a different shift moves the [sampler off them]\"; the same reasoning applies to any Turbo-trained adapter in the chain.", + "Reference/edit LoRAs are designed to be stacked under ordinary subject and style LoRAs - the Identity Edit card states \"Composes with your LoRAs: character/body/style LoRAs stack on top and steer the prior\" and recommends a stacked subject LoRA when unusual subjects drift toward the base prior.", + "Krea 2 LoRAs cannot be chained with LoRAs for any other base model. Many CivitAI pages carry multi-model titles like \"(Krea-2 + ZIT)\" or \"[Flux | ZIB | Krea2]\" - those are separate files for separate architectures shipped under one model page, not a combined adapter.", + "To combine two full Krea 2 checkpoints rather than adapters, ComfyUI ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per-section weights for first., tmlp., txtmlp., tproj., txtfusion.projector, txtfusion.layerwise_blocks.0-1, txtfusion.refiner_blocks.0-1, blocks.0-27 and last.", + "The author never states which weight family this targets. Silent degradation, never a hard error, between Raw and Turbo. The two share one architecture - ai-toolkit's config is commented \"The reference 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)\" - so every LoRA tensor shape matches on both and ComfyUI/diffuser" + ] + }, + "pros": [ + "Widely adopted — 29,303 downloads. [source: CivitAI stats]", + "Positively received — 3,234 thumbs-up, 1 down. [source: CivitAI stats]", + "Explicit trigger word(s) — wearing a green slavekini top with burgundy loin cloth, gold choker, gold armlet, and a single wrist cuff, hair up in a gold hair ornament, with hair in braid over shoulder with gold hair tie — so the effect can be turned on and off from the prompt. [source: model metadata]", + "Compact at 218 MB — cheap to stack with other LoRAs. [source: file size]", + "Author published concrete recommended settings. [source: description]" + ], + "cons": [ + "Needs a reduced strength; the default 1.0 is likely too strong. [source: description]", + "Author references NSFW output — review before using in shared work. [source: description]", + "Author never states which generation mode it was trained for; test against your own workflow before relying on it. [source: description]" + ], + "file": { + "name": "Slavekini_Krea2_epoch_7.safetensors", + "size_bytes": 228588904, + "size_mb": 218.0, + "sha256": "ff593bf2f0b67a8101e3adb894ec94b5d16dfb3b7dd69fbec9df488ca772d33e" + }, + "stats": { + "downloads": 29303, + "thumbs_up": 3234, + "thumbs_down": 1 + }, + "popularity": { + "downloads": 29303, + "thumbs_up": 3234, + "thumbs_down": 1, + "comments": 0, + "approval": "100% positive (3234/3235)", + "age_days": 45.2, + "downloads_per_day": 26.4, + "rank_by_downloads": 20, + "rank_by_momentum": 84, + "of_total": 150, + "adoption_tier": "High", + "momentum": "Long tail" + }, + "selection_score": 68511.0, + "selection_reasons": [ + "29,303 downloads", + "3,234 thumbs-up", + "documented triggers + usage notes", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "1152x1792", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "1024x1536", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "1120x1664", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "1152x1792", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "1152x1792", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "1024x1536", + "prompt": "" + } + ], + "tags": [ + "swimsuit", + "anime", + "sexy", + "realistic", + "woman", + "clothing", + "photorealistic", + "star wars", + "fantasy" + ], + "scraped_at": "2026-09-05T00:11:38+00:00", + "path": "krea-2/slavekini-aka-slave-leia-outfit-clothing-krea2-anima-il" +} diff --git a/the-space-marines-warhammer-40k-flux-pony-illustrious-anima/Krea2UltraM40k.safetensors b/the-space-marines-warhammer-40k-flux-pony-illustrious-anima/Krea2UltraM40k.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..a7e443d46916542a50c051597f9d33b1498b2b19 --- /dev/null +++ b/the-space-marines-warhammer-40k-flux-pony-illustrious-anima/Krea2UltraM40k.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:507e7d952e044362583b3525e537596f5d3c9e12096a3be15f5cd62ed3c5c1ff +size 228588882 diff --git a/the-space-marines-warhammer-40k-flux-pony-illustrious-anima/example_images/01_image.jpeg b/the-space-marines-warhammer-40k-flux-pony-illustrious-anima/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..965ae8c1eeb17768db9a405b45f758b20d6ad684 --- /dev/null +++ b/the-space-marines-warhammer-40k-flux-pony-illustrious-anima/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e7c023b8f9745f382623684024f613a31c720d155b74591793142fab3d056a16 +size 1983067 diff --git a/the-space-marines-warhammer-40k-flux-pony-illustrious-anima/example_images/02_image.jpeg b/the-space-marines-warhammer-40k-flux-pony-illustrious-anima/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..87fc9ebcb672bcb3b5b29fe00301505a82fb4b00 --- /dev/null +++ b/the-space-marines-warhammer-40k-flux-pony-illustrious-anima/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d6d1b62b73be26f5c7b9880ac14b495bab2a1de3415aad0d81058c6ef82b6db0 +size 1234319 diff --git a/the-space-marines-warhammer-40k-flux-pony-illustrious-anima/example_images/03_image.jpeg b/the-space-marines-warhammer-40k-flux-pony-illustrious-anima/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..ee9e207af42de7ca8dda1bd1f4c9f573cdbc3605 --- /dev/null +++ b/the-space-marines-warhammer-40k-flux-pony-illustrious-anima/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ceb65bc12515cf3d64d9d08528c5f53bcb7168b189dace5324172967aea45dd +size 1362952 diff --git a/the-space-marines-warhammer-40k-flux-pony-illustrious-anima/example_images/04_image.jpeg b/the-space-marines-warhammer-40k-flux-pony-illustrious-anima/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..5b677606e958278c4c0d00d1eea139cd072a13a2 --- /dev/null +++ b/the-space-marines-warhammer-40k-flux-pony-illustrious-anima/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5ee21f821f76564deb6388b70a8547715b6da54b7cf01fbf1dab8dc60e696a0 +size 1353302 diff --git a/the-space-marines-warhammer-40k-flux-pony-illustrious-anima/example_images/05_image.jpeg b/the-space-marines-warhammer-40k-flux-pony-illustrious-anima/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..9539b786b80e4f7ce0ebeaaddde8a63fcfadc6a4 --- /dev/null +++ b/the-space-marines-warhammer-40k-flux-pony-illustrious-anima/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ba16189f6888a249e07aadf393ec4348f2e8edf452434793dd8c9b246f1d700a +size 2799652 diff --git a/the-space-marines-warhammer-40k-flux-pony-illustrious-anima/example_images/06_image.jpeg b/the-space-marines-warhammer-40k-flux-pony-illustrious-anima/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..7039a1992bc961a5d71e8ec99891a106345f99a9 --- /dev/null +++ b/the-space-marines-warhammer-40k-flux-pony-illustrious-anima/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:03f8f59f529c0f6e7e132d368afe07ef0fd92dd7cc4cc1bb130671039413bd87 +size 227057 diff --git a/the-space-marines-warhammer-40k-flux-pony-illustrious-anima/info.txt b/the-space-marines-warhammer-40k-flux-pony-illustrious-anima/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..2db76b817f1f0bbc944d51c2b9fdcc2fe69e7937 --- /dev/null +++ b/the-space-marines-warhammer-40k-flux-pony-illustrious-anima/info.txt @@ -0,0 +1,346 @@ +═══════════════════════════════════════════════════════════════════════════════ + THE SPACE MARINES WARHAMMER 40K | FLUX + PONY + ILLUSTRIOUS + ANIMA + KREA2 +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Character / subject + Author VelvetS + Source https://civitai.com/models/632900 + Provenance CivitAI · model 632900 · version 3145005 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File Krea2UltraM40k.safetensors + Version Ultramarines Krea2 + Size 218.0 MB + Format SafeTensor / unknown precision + SHA256 507e7d952e044362583b3525e537596f5d3c9e12096a3be15f5cd62ed3c5c1ff + Published 2026-07-19 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + › "UltraM40k" + +─────────────────────────────────────────────────────────────────────────────── + RECOMMENDED SETTINGS +─────────────────────────────────────────────────────────────────────────────── + The author did not publish specific settings. + No author settings were published. Community baseline: steps Turbo: 8 + (official README and HF card CLI example; the ComfyUI template ships 8). + Raw: 52 (official README). The krea-2 CLI's generic default is 28. + Community Turbo LoRAs commonly suggest 8-12 (e.g. the Identity Edit card + recommends Turbo at 8-12, or Raw at 20 for removals). A community 4-step + LoRA on Turbo makes 4 valid; its author calls 2 steps an out-of-spec + preview mode., cfg Turbo: CFG disabled - guidance_scale 0.0 in diffusers + and --cfg 0.0 in the krea-2 CLI, which is cfg 1.0 in ComfyUI's KSampler + (the official template ships cfg 1). These are the same thing; do not set + ComfyUI cfg to 0. Raw: CFG 3.5 (official README; CLI generic default 4.5). + Negative prompting on Turbo is effectively unused - the official templates + route ConditioningZeroOut into the negative slot., sampler ComfyUI: + KSampler with sampler euler, scheduler simple, denoise 1.0 (verified in + image_krea2_turbo_t2i.json: [seed, 'randomize', 8, 1, 'euler', 'simple', + 1]). Timestep shift mu = 1.15 for Turbo; ComfyUI bakes this in via + Krea2.sampling_settings = {"multiplier": 1.0, "shift": 1.15}, so the plain + t2i template needs no ModelSampling node, while the style-reference + template adds ModelSamplingFlux at max_shift 1.15 / base_shift 0.5. The CLI + exposes --mu 1.15 for Turbo; Raw uses a resolution-derived mu interpolated + between --y1 0.5 and --y2 1.15 (ai-toolkit mirrors this as base_shift 0.5 / + max_shift 1.15 over image_seq_len 256-6400, exponential time shift)., + resolution 1024x1024 default. Turbo: 1K-2K (the official CLI example + generates 2048x2048; ComfyUI's ResolutionSelector defaults to 1 megapixel, + set 2.0 for 2K). Raw: "trained to generate upto 1k resolution". + Width/height are padded up to a multiple of 16., lora_strength 1.0 for + Krea's nine official style LoRAs (docs.comfy.org lists Recommended Strength + 1.0 for every one), though the shipped t2i template's LoraLoaderModelOnly + is set to 0.8. The style-reference LoRA is used at 1.0 in its template. The + 4-step distill LoRA documents 1.0 as its trained point, 1.0-1.5 as coherent + extrapolation and above ~1.5 as speckle artefacting. Community subject + LoRAs are typically recommended around 0.8-1.2 and dropped toward 0.6-0.9 + when several are stacked.. + +─────────────────────────────────────────────────────────────────────────────── + ADOPTION & POPULARITY +─────────────────────────────────────────────────────────────────────────────── + Downloads 33,547 + Rating 100% positive (1872/1872) + Comments 0 + Published 47 days ago + Download rate 10.3/day since release + + Adoption rank #15 of 150 in this collection (Top tier) + Momentum rank #138 of 150 by download rate (Long tail) + + Adoption tier is the percentile of total downloads within this collection; + momentum is the percentile of downloads-per-day since release. Momentum is + ranked rather than measured against a fixed rate, because this base model + is itself new and every LoRA looks fast on an absolute scale. Both numbers + are CivitAI's, read on the scrape date at the foot of this file. + +─────────────────────────────────────────────────────────────────────────────── + CHAINING / STACKING +─────────────────────────────────────────────────────────────────────────────── + Stacks with others not stated + + The author gave no chaining guidance. The rules below are the + general Krea 2 ones, not anything specific to this LoRA. + + General Krea 2 rules that apply here (see REFERENCE.txt): + + • The most important rule is not about chaining but about which + checkpoint you chain onto: the official README says in bold "TRAIN on + Raw and RUN on Turbo", and "We highly recommend using RAW for training + LoRAs and applying them on Turbo for inference." LoRAs trained on Raw + are stated to work well on Turbo. + • Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 + LoRAs are DiT-only, so there is no CLIP side to chain and the text- + encoder branch stays untouched however many LoRAs you add. + • Krea 2 LoRAs stack in practice and the community treats multi-LoRA as + routine - style + character, style + slider, realism + character. + Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one + writes "I recently stacked over 20 of my own LoRAs on a single + generation and the output still came out great." + • Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge + on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is + loaded. The CivitAI 12GB training guide also notes rank affects + composability: "rank 8 produced a lighter style influence that was + easier to stack with other LoRAs, rank 16 provided a useful balance + between concept strength and composability, and rank 32 was better + suited to a precise object or concept intended to take priority over + other stacked LoRAs." + • Accelerator LoRAs are position-sensitive in intent, not in node order: + a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a + sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to- + Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is + explicitly recommended by the 4-step LoRA's author. + • Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI + (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card + says "Keep mu = 1.15. The training targets are anchored to that grid; a + different shift moves the [sampler off them]"; the same reasoning + applies to any Turbo-trained adapter in the chain. + • Reference/edit LoRAs are designed to be stacked under ordinary subject + and style LoRAs - the Identity Edit card states "Composes with your + LoRAs: character/body/style LoRAs stack on top and steer the prior" and + recommends a stacked subject LoRA when unusual subjects drift toward + the base prior. + • Krea 2 LoRAs cannot be chained with LoRAs for any other base model. + Many CivitAI pages carry multi-model titles like "(Krea-2 + ZIT)" or + "[Flux | ZIB | Krea2]" - those are separate files for separate + architectures shipped under one model page, not a combined adapter. + • To combine two full Krea 2 checkpoints rather than adapters, ComfyUI + ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per- + section weights for first., tmlp., txtmlp., tproj., + txtfusion.projector, txtfusion.layerwise_blocks.0-1, + txtfusion.refiner_blocks.0-1, blocks.0-27 and last. + • The author never states which weight family this targets. Silent + degradation, never a hard error, between Raw and Turbo. The two share + one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" + - so every LoRA tensor shape matches on both and ComfyUI/diffuser + +─────────────────────────────────────────────────────────────────────────────── + PROS +─────────────────────────────────────────────────────────────────────────────── + + Widely adopted — 33,547 downloads. [source: CivitAI stats] + + Positively received — 1,872 thumbs-up with no down-votes. [source: + CivitAI stats] + + Explicit trigger word(s) — UltraM40k — so the effect can be turned on + and off from the prompt. [source: model metadata] + + Compact at 218 MB — cheap to stack with other LoRAs. [source: file + size] + +─────────────────────────────────────────────────────────────────────────────── + CONS & CAVEATS +─────────────────────────────────────────────────────────────────────────────── + - Author never states which generation mode it was trained for; test + against your own workflow before relying on it. [source: description] + +─────────────────────────────────────────────────────────────────────────────── + COMPATIBILITY WARNING +─────────────────────────────────────────────────────────────────────────────── + Silent degradation, never a hard error, between Raw and Turbo. The two + share one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" - so + every LoRA tensor shape matches on both and ComfyUI/diffusers will load an + adapter against either without complaint. The lvladikov 4-step LoRA card + states this for the reverse direction: "Every layer it targets also exists + in Krea 2 Raw, so it will load there without complaint - but that is a side + effect of the shared architecture, not a supported mode", and reports + mixed, subject-dependent results with malformed subjects (duplicated heads, + fused limbs) at low step counts, needing 14+ steps with Raw's CFG on before + output is coherent. The intended direction (train on Raw, run on Turbo) is + officially blessed. The more common practical failure is a settings + mismatch rather than a weight mismatch: running a Turbo-targeted LoRA at + Raw's 52 steps + CFG 3.5, or leaving CFG enabled on Turbo, degrades output + with no warning. Loading a Krea 2 LoRA against a different model family + entirely (Flux, Qwen-Image, Z-Image) is also not an error - ComfyUI logs + "lora key not loaded: {key}" per key and generates as if no LoRA were + applied. + +─────────────────────────────────────────────────────────────────────────────── + AUTHOR'S DESCRIPTION +─────────────────────────────────────────────────────────────────────────────── + This LoRA is for Pony Diffusion XL, illustrious and FLUX.1 [dev], it's made + to give your character the Space Marines power armor from the world of + Warhammer 40K. + + Deathwatch "Suffer Not The Alien to Live!" Flux: Trigger word: DeathW40K + Works on: 0.6-1 weights, I prefer 0.9 + Pony: Trigger word: DeathW40K + Works on: 0.6-1 weights + + Thousand Sons "All is dust!" Flux: Trigger word: ThouS40k + Works on: 0.6-1 weights, I prefer 0.9 + Pony: Trigger word: ThouS40k + Works on: 0.6-1 weights + + Grey Knights "We are the hammer!" Flux: Trigger word: GreyK40k + Works on: 0.6-1 weights, I prefer 0.9 + Pony: Trigger word: GreyK40k + Works on: 0.6-1 weights + + Dark Angels "For the Lion!" Flux: Trigger word: DarkA40k + Works on: 0.6-1 weights, I prefer 0.9 + Pony: Trigger word: DarkA40k + Works on: 0.6-1 weights + + Raven Guard "Victorus Aut Mortis!" Flux: Trigger word: RavenG40k + Works on: 0.6-1 weights, I prefer 0.9 + Pony: Trigger word: RavenG40k + Works on: 0.6-1 weights + + Blood Angels "By the Blood of Sanguinius!" Flux: Trigger word: BloodA40K + Works on: 0.6-1 weights, I prefer 0.9 + Pony: Trigger word: BloodA40K + Works on: 0.6-1 weights + + Alpha Legion "Hydra Dominatus!" Pony: Trigger word: AlphaL40k + Works on: 0.6-1 weights + (Alpha Legion flux version didn't look as good as I wanted so it might take + a while) + + Space Wolves "For Russ and the Allfather!" Flux: Trigger word: SpaceW40K + Works on: 0.6-1 weights, I prefer 0.9 + Pony: Trigger word: SpaceW40K + Works on: 0.6-1 weights + + Night Lords "Ave Dominus Nox!" Flux: Trigger word: NightL40k + Works on: 0.6-1 weights, I prefer 0.9 + Pony: Trigger word: NightL40k + Works on: 0.6-1 weights + + White Scars "For the Khan and the Emperor!" Flux: Trigger word: WhiteS40k + Works on: 0.6-1 weights, I prefer 0.9 + Pony: Trigger word: WhiteS40k + Works on: 0.6-1 weights + + Imperial Fists "Primarch-Progenitor, to your glory and the glory of him on + earth!" Flux: Trigger word: ImperialF40k + Works on: 0.6-1 weights, I prefer 0.9 + Pony: Trigger word: ImperialF40k + Works on: 0.6-1 weights + + Black Templars "No Pity! No Remorse! No Fear!" Flux: Trigger word: + BlackT40k + Works on: 0.6-1 weights, I prefer 0.9 + Pony: Trigger word: BlackT40k + Works on: 0.6-1 weights + + Ultramarines "Courage and Honour!" Flux: Trigger word: UltraM40k + Works on: 0.6-1 weights, I prefer 0.9 + Pony: Trigger word: UltraM40k + Works on: 0.6-1 weights + + More Space Marine Chapters soon + Can't wait to see what you guys make + enjoy :) + +─────────────────────────────────────────────────────────────────────────────── + EXAMPLE MEDIA +─────────────────────────────────────────────────────────────────────────────── + 6 preview(s) in ./example_images/ — community results for + this LoRA. Videos keep a matching _poster.jpg still frame. + + No prompts are recorded below: CivitAI's public API flags these posts as + having generation metadata but does not return it, so the prompts cannot be + scraped. Open the model page to read them. + + 01_image.jpeg (image, 960x1280) + 02_image.jpeg (image, 864x1080) + 03_image.jpeg (image, 864x1080) + 04_image.jpeg (image, 864x1080) + 05_image.jpeg (image, 1072x1616) + 06_image.jpeg (image, 768x1376) + +─────────────────────────────────────────────────────────────────────────────── + HOW TO USE (COMFYUI) +─────────────────────────────────────────────────────────────────────────────── + 1. Update ComfyUI to at least v0.26.0 (released 2026-06-23), which added + local Krea2 support. For reference-image / style-reference / edit LoRAs you + need at least v0.29.0 (2026-07-29), which added regular and timestep-zero + reference images for Krea 2. ModelMergeKrea2 arrived in v0.27.0 + (2026-06-30). + + 2. Put the base stack in place (files from huggingface.co/Comfy-Org/Krea-2, + which is not gated): + - ComfyUI/models/diffusion_models/krea2_turbo_fp8_scaled.safetensors + (documented recommended default; the repo also holds krea2_turbo_bf16, + _int8_convrot, _mxfp8, _nvfp4 and krea2_raw_bf16 / _fp8_scaled / + _int8_convrot) + - ComfyUI/models/text_encoders/qwen3vl_4b_fp8_scaled.safetensors (bf16 + also available) + - ComfyUI/models/vae/qwen_image_vae.safetensors + + 3. Put the LoRA .safetensors in ComfyUI/models/loras/ - flat, no per-model + subfolder required. + + 4. Build or load the graph. The official image_krea2_turbo_t2i.json + template is: UNETLoader (krea2_turbo_fp8_scaled.safetensors) -> + LoraLoaderModelOnly -> KSampler; CLIPLoader + (qwen3vl_4b_fp8_scaled.safetensors, type "krea2") -> text encode -> + KSampler positive, with ConditioningZeroOut into negative; VAELoader + (qwen_image_vae.safetensors) -> VAEDecode -> SaveImage; EmptyLatentImage + 1024x1024 driven by ResolutionSelector. An int8 sibling template + (image_krea2_turbo_t2i_int8.json) also ships. + + 5. Load the LoRA with LoraLoaderModelOnly, not LoraLoader - both official + templates do. Krea 2 LoRAs patch the DiT only; note that the DiT's own + text_fusion / txtfusion blocks are a normal LoRA target (the official style + LoRAs patch them), but the Qwen3-VL text encoder itself is not. Set + strength 1.0 for Krea's official style LoRAs (the shipped template uses + 0.8) and add the LoRA's trigger word to the prompt. + + 6. Sampler settings: steps 8, cfg 1, sampler euler, scheduler simple, + denoise 1.0. Do not add a ModelSampling node for plain t2i - ComfyUI + already applies shift 1.15 from Krea2.sampling_settings. + + 7. For style reference, use the dedicated template: UNETLoader + krea2_turbo_int8_convrot.safetensors, LoraLoaderModelOnly + krea2_style_reference.safetensors at 1.0, ModelSamplingFlux (1.15, 0.5), + TextEncodeQwenImageEditPlus fed by LoadImage, + FluxKontextMultiReferenceLatentMethod set to index_timestep_zero, and + SamplerCustomAdvanced with KSamplerSelect euler + BasicScheduler simple/8 + + CFGGuider cfg 1. + + 8. For instruction editing (Krea 2 Identity Edit), stock nodes are not + enough: install the ComfyUI-Krea2Edit node pack, which ships its own + workflows, and run Turbo at 8-12 steps / cfg 1 (or Raw at 20 steps / cfg 3 + for removals) with LoRA strength 1.0. + + 9. Low-VRAM alternative: install city96's ComfyUI-GGUF and use a community + GGUF build (realrebelai/KREA-2_GGUFs, quantized from Turbo; + molbal/krea2-gguf, quantized from Raw) with the Unet Loader (GGUF) node in + place of UNETLoader. There is no official GGUF release. LoRA loading is + unchanged. + +─────────────────────────────────────────────────────────────────────────────── + Scraped by lorakit on 2026-09-05 00:12 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/the-space-marines-warhammer-40k-flux-pony-illustrious-anima/metadata.json b/the-space-marines-warhammer-40k-flux-pony-illustrious-anima/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..f9bdfc7c14b64d354468c047823c52237bafbff8 --- /dev/null +++ b/the-space-marines-warhammer-40k-flux-pony-illustrious-anima/metadata.json @@ -0,0 +1,139 @@ +{ + "name": "The Space Marines Warhammer 40K | Flux + Pony + illustrious + Anima + Krea2", + "slug": "the-space-marines-warhammer-40k-flux-pony-illustrious-anima", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 632900, + "version_id": 3145005, + "version_name": "Ultramarines Krea2", + "source": "civitai", + "url": "https://civitai.com/models/632900", + "repo_id": null, + "file_path": null, + "author": "VelvetS", + "category": { + "key": "character", + "label": "Character / subject" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [ + "UltraM40k" + ], + "recommended_settings": {}, + "chaining": { + "author": { + "stacks": "not stated", + "notes": "", + "avoid": [], + "requires": [], + "order": "", + "strength": "", + "stated": false + }, + "general_rules": [ + "The most important rule is not about chaining but about which checkpoint you chain onto: the official README says in bold \"TRAIN on Raw and RUN on Turbo\", and \"We highly recommend using RAW for training LoRAs and applying them on Turbo for inference.\" LoRAs trained on Raw are stated to work well on Turbo.", + "Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 LoRAs are DiT-only, so there is no CLIP side to chain and the text-encoder branch stays untouched however many LoRAs you add.", + "Krea 2 LoRAs stack in practice and the community treats multi-LoRA as routine - style + character, style + slider, realism + character. Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one writes \"I recently stacked over 20 of my own LoRAs on a single generation and the output still came out great.\"", + "Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is loaded. The CivitAI 12GB training guide also notes rank affects composability: \"rank 8 produced a lighter style influence that was easier to stack with other LoRAs, rank 16 provided a useful balance between concept strength and composability, and rank 32 was better suited to a precise object or concept intended to take priority over other stacked LoRAs.\"", + "Accelerator LoRAs are position-sensitive in intent, not in node order: a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to-Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is explicitly recommended by the 4-step LoRA's author.", + "Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card says \"Keep mu = 1.15. The training targets are anchored to that grid; a different shift moves the [sampler off them]\"; the same reasoning applies to any Turbo-trained adapter in the chain.", + "Reference/edit LoRAs are designed to be stacked under ordinary subject and style LoRAs - the Identity Edit card states \"Composes with your LoRAs: character/body/style LoRAs stack on top and steer the prior\" and recommends a stacked subject LoRA when unusual subjects drift toward the base prior.", + "Krea 2 LoRAs cannot be chained with LoRAs for any other base model. Many CivitAI pages carry multi-model titles like \"(Krea-2 + ZIT)\" or \"[Flux | ZIB | Krea2]\" - those are separate files for separate architectures shipped under one model page, not a combined adapter.", + "To combine two full Krea 2 checkpoints rather than adapters, ComfyUI ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per-section weights for first., tmlp., txtmlp., tproj., txtfusion.projector, txtfusion.layerwise_blocks.0-1, txtfusion.refiner_blocks.0-1, blocks.0-27 and last.", + "The author never states which weight family this targets. Silent degradation, never a hard error, between Raw and Turbo. The two share one architecture - ai-toolkit's config is commented \"The reference 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)\" - so every LoRA tensor shape matches on both and ComfyUI/diffuser" + ] + }, + "pros": [ + "Widely adopted — 33,547 downloads. [source: CivitAI stats]", + "Positively received — 1,872 thumbs-up with no down-votes. [source: CivitAI stats]", + "Explicit trigger word(s) — UltraM40k — so the effect can be turned on and off from the prompt. [source: model metadata]", + "Compact at 218 MB — cheap to stack with other LoRAs. [source: file size]" + ], + "cons": [ + "Author never states which generation mode it was trained for; test against your own workflow before relying on it. [source: description]" + ], + "file": { + "name": "Krea2UltraM40k.safetensors", + "size_bytes": 228588882, + "size_mb": 218.0, + "sha256": "507e7d952e044362583b3525e537596f5d3c9e12096a3be15f5cd62ed3c5c1ff" + }, + "stats": { + "downloads": 33547, + "thumbs_up": 1872, + "thumbs_down": 0 + }, + "popularity": { + "downloads": 33547, + "thumbs_up": 1872, + "thumbs_down": 0, + "comments": 0, + "approval": "100% positive (1872/1872)", + "age_days": 47.3, + "downloads_per_day": 10.31, + "rank_by_downloads": 15, + "rank_by_momentum": 138, + "of_total": 150, + "adoption_tier": "Top tier", + "momentum": "Long tail" + }, + "selection_score": 56411.0, + "selection_reasons": [ + "33,547 downloads", + "1,872 thumbs-up", + "documented triggers + usage notes", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "960x1280", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "864x1080", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "864x1080", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "864x1080", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "1072x1616", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "768x1376", + "prompt": "" + } + ], + "tags": [ + "armor", + "woman", + "40k", + "man", + "space marine", + "clothing", + "warhammer", + "ultramarines" + ], + "scraped_at": "2026-09-05T00:12:01+00:00", + "path": "krea-2/the-space-marines-warhammer-40k-flux-pony-illustrious-anima" +} diff --git a/ultrareal-krea2-klein9b/example_images/01_image.jpeg b/ultrareal-krea2-klein9b/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..28d9c7e419db365f5256fff224d8a0f68c5cbcf0 --- /dev/null +++ b/ultrareal-krea2-klein9b/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b44ae1cf6c1b2c2cbe4b25996772bfd105abdbfca6718adb61ad35e3f77f1de4 +size 348270 diff --git a/ultrareal-krea2-klein9b/example_images/02_image.jpeg b/ultrareal-krea2-klein9b/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..2f3efe70716c29150f22844b12f97cd778ed6c8f --- /dev/null +++ b/ultrareal-krea2-klein9b/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8dfd089cc63b86f53a804ab3bba2af8ccc2e794d2c09d0f4e70fb2b472e47f9d +size 1513486 diff --git a/ultrareal-krea2-klein9b/example_images/03_image.jpeg b/ultrareal-krea2-klein9b/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..3090dd9b09bb471c4bb07d2ddeb431cf81d03258 --- /dev/null +++ b/ultrareal-krea2-klein9b/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eac3fa5e93bfadfdda37bd18ebe50aaa6f4bee2559139cc4fc42e0270f340647 +size 335801 diff --git a/ultrareal-krea2-klein9b/example_images/04_image.jpeg b/ultrareal-krea2-klein9b/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..2fa799621261a2acd62fc973633e88863b685356 --- /dev/null +++ b/ultrareal-krea2-klein9b/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cc8d7a37b5f74c52b942f4fc1116781727770ce2f667c7830d4cb5a7331c30cf +size 2265050 diff --git a/ultrareal-krea2-klein9b/example_images/05_image.jpeg b/ultrareal-krea2-klein9b/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..78d65d2b86e4457d404b03dbbbee76f9d183cd20 --- /dev/null +++ b/ultrareal-krea2-klein9b/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2337ea909289407a57ca5a8fa997c3333a425a9b8a91ee49f0139f6f695d9cde +size 1991753 diff --git a/ultrareal-krea2-klein9b/example_images/06_image.jpeg b/ultrareal-krea2-klein9b/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..c077c43265c320e54c6d1aa23e401910c99ed7db --- /dev/null +++ b/ultrareal-krea2-klein9b/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0490dbb3c0d7daf6b8bacb262ed075e6e2e714a025353c581e15c99d727bab48 +size 344699 diff --git a/ultrareal-krea2-klein9b/info.txt b/ultrareal-krea2-klein9b/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..9311734018a79e8bb1ef6caa403b02f42eaac3e3 --- /dev/null +++ b/ultrareal-krea2-klein9b/info.txt @@ -0,0 +1,297 @@ +═══════════════════════════════════════════════════════════════════════════════ + ULTRAREAL - KREA2, KLEIN9B +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Photorealism + Author vizsumit + Source https://civitai.com/models/2462105 + Provenance CivitAI · model 2462105 · version 3182356 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File ultra_real_krea2_v2.safetensors + Version KR2_V2 + Size 218.0 MB + Format SafeTensor / unknown precision + SHA256 233a8b1df4b3387f9f2bedaa2099d0e14cc946d1105f732de2a8600310b86f07 + Published 2026-07-30 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + No trigger word. The LoRA applies as soon as it is loaded — + strength is the only way to dial the effect up or down. + +─────────────────────────────────────────────────────────────────────────────── + RECOMMENDED SETTINGS +─────────────────────────────────────────────────────────────────────────────── + Taken verbatim from the author's description: + + LoRA strength 0.6–0.8 + +─────────────────────────────────────────────────────────────────────────────── + ADOPTION & POPULARITY +─────────────────────────────────────────────────────────────────────────────── + Downloads 36,247 + Rating 100% positive (1104/1104) + Comments 0 + Published 36 days ago + Download rate 196.5/day since release + + Adoption rank #14 of 150 in this collection (Top tier) + Momentum rank #11 of 150 by download rate (Hot · fresh release) + + Adoption tier is the percentile of total downloads within this collection; + momentum is the percentile of downloads-per-day since release. Momentum is + ranked rather than measured against a fixed rate, because this base model + is itself new and every LoRA looks fast on an absolute scale. Both numbers + are CivitAI's, read on the scrape date at the foot of this file. + +─────────────────────────────────────────────────────────────────────────────── + CHAINING / STACKING +─────────────────────────────────────────────────────────────────────────────── + Stacks with others not stated + + The author gave no chaining guidance. The rules below are the + general Krea 2 ones, not anything specific to this LoRA. + + General Krea 2 rules that apply here (see REFERENCE.txt): + + • The most important rule is not about chaining but about which + checkpoint you chain onto: the official README says in bold "TRAIN on + Raw and RUN on Turbo", and "We highly recommend using RAW for training + LoRAs and applying them on Turbo for inference." LoRAs trained on Raw + are stated to work well on Turbo. + • Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 + LoRAs are DiT-only, so there is no CLIP side to chain and the text- + encoder branch stays untouched however many LoRAs you add. + • Krea 2 LoRAs stack in practice and the community treats multi-LoRA as + routine - style + character, style + slider, realism + character. + Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one + writes "I recently stacked over 20 of my own LoRAs on a single + generation and the output still came out great." + • Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge + on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is + loaded. The CivitAI 12GB training guide also notes rank affects + composability: "rank 8 produced a lighter style influence that was + easier to stack with other LoRAs, rank 16 provided a useful balance + between concept strength and composability, and rank 32 was better + suited to a precise object or concept intended to take priority over + other stacked LoRAs." + • Accelerator LoRAs are position-sensitive in intent, not in node order: + a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a + sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to- + Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is + explicitly recommended by the 4-step LoRA's author. + • Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI + (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card + says "Keep mu = 1.15. The training targets are anchored to that grid; a + different shift moves the [sampler off them]"; the same reasoning + applies to any Turbo-trained adapter in the chain. + • Reference/edit LoRAs are designed to be stacked under ordinary subject + and style LoRAs - the Identity Edit card states "Composes with your + LoRAs: character/body/style LoRAs stack on top and steer the prior" and + recommends a stacked subject LoRA when unusual subjects drift toward + the base prior. + • Krea 2 LoRAs cannot be chained with LoRAs for any other base model. + Many CivitAI pages carry multi-model titles like "(Krea-2 + ZIT)" or + "[Flux | ZIB | Krea2]" - those are separate files for separate + architectures shipped under one model page, not a combined adapter. + • To combine two full Krea 2 checkpoints rather than adapters, ComfyUI + ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per- + section weights for first., tmlp., txtmlp., tproj., + txtfusion.projector, txtfusion.layerwise_blocks.0-1, + txtfusion.refiner_blocks.0-1, blocks.0-27 and last. + • The author never states which weight family this targets. Silent + degradation, never a hard error, between Raw and Turbo. The two share + one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" + - so every LoRA tensor shape matches on both and ComfyUI/diffuser + +─────────────────────────────────────────────────────────────────────────────── + PROS +─────────────────────────────────────────────────────────────────────────────── + + Widely adopted — 36,247 downloads. [source: CivitAI stats] + + Positively received — 1,104 thumbs-up with no down-votes. [source: + CivitAI stats] + + Compact at 218 MB — cheap to stack with other LoRAs. [source: file + size] + + Author published concrete recommended settings. [source: description] + +─────────────────────────────────────────────────────────────────────────────── + CONS & CAVEATS +─────────────────────────────────────────────────────────────────────────────── + - Author references NSFW output — review before using in shared work. + [source: description] + - No trigger word — the LoRA is always on once loaded, so strength is the + only control. [source: model metadata] + - Author never states which generation mode it was trained for; test + against your own workflow before relying on it. [source: description] + +─────────────────────────────────────────────────────────────────────────────── + COMPATIBILITY WARNING +─────────────────────────────────────────────────────────────────────────────── + Silent degradation, never a hard error, between Raw and Turbo. The two + share one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" - so + every LoRA tensor shape matches on both and ComfyUI/diffusers will load an + adapter against either without complaint. The lvladikov 4-step LoRA card + states this for the reverse direction: "Every layer it targets also exists + in Krea 2 Raw, so it will load there without complaint - but that is a side + effect of the shared architecture, not a supported mode", and reports + mixed, subject-dependent results with malformed subjects (duplicated heads, + fused limbs) at low step counts, needing 14+ steps with Raw's CFG on before + output is coherent. The intended direction (train on Raw, run on Turbo) is + officially blessed. The more common practical failure is a settings + mismatch rather than a weight mismatch: running a Turbo-targeted LoRA at + Raw's 52 steps + CFG 3.5, or leaving CFG enabled on Turbo, degrades output + with no warning. Loading a Krea 2 LoRA against a different model family + entirely (Flux, Qwen-Image, Z-Image) is also not an error - ComfyUI logs + "lora key not loaded: {key}" per key and generates as if no LoRA were + applied. + +─────────────────────────────────────────────────────────────────────────────── + AUTHOR'S DESCRIPTION +─────────────────────────────────────────────────────────────────────────────── + LoRAs for reducing the soft & smooth AI look and adds more texture and + details to images. It works especially well for close-ups and medium shots + where clothing details and skin texture plays an important role. + Run this exact workflow instantly on FloyoAI, no installation, no setup: 👉 + https://www.floyo.ai/workflows/krea2-ultra-real-o7ys4feqdhpm Use code + vizsumit to get 1 month free (first 100 users only) + Version Comparison + Krea_V2 is trained on high-qulality Professional Photography images with + wide variety like portraits, lifestyle, fashion, wildlife, sports fitness, + architecutre, nature and more. Since dataset had professional photography + as focus, it tends to produce studio quality lighting and textures. So + prefer this for professionalism. + Krea_V1 is trained on high-qulality SFW and NSFW images so it can handle + both. Besides making images more detailed it also reduces asian face bias . + V4 is trained on more refined and 40% nsfw dataset, so besides better + quality it has NSFW detailing capabilities too. + V3 is trained on professional 4K quality images with higher steps. It + applies very subtle skin texture which looks more realistic, without + producing freckles and burnes. + V2 for more natural looking skin texture than V1, can produce freckles and + moles for some time. It is good at preserving skin tone and lighting also. + V1 tends to produce overdone skin texture like more pores and freckles, and + it can change lighting and skin tone also. + 🖼️ Generation Settings + Weight: 0.6-0.8 + Krea2 Trigger: no trigger needed. + V4 Trigger: This is a high-quality photograph of a (your subject...) + V3 Trigger: A high-quality photograph of a (your subject...) + V2 Trigger: A high-quality medium close-up photo with realistic skin + texture featuring a (your subject...) + V1 Trigger: This is a high-quality photo with realistic skin texture and + details of a (your subject...) + 🛠️ Editing Settings (Klein9b) + V4 Weight: 0.55, Prompt: Make this photo high-quality. Preserve subject's + face, expression, figure and pose. Preserve overall composition of this + photo. + V3 Weight: 0.6, Prompt: Make this photo high-quality. Preserve subject's + face, expression, figure and pose. Preserve overall composition of this + photo. + V2 Weight: 0.6, Prompt: Make this photo high-quality with natural skin + texture. Preserve subject's facial features, expression, figure and pose. + Preserve overall composition of this photo. + V1 Weight: 0.5, Prompt: Make this photo high-quality featuring realstic + skin texture. Preserve subject's facial features, expression, figure and + pose. Preserve overall composition of this photo. + Feel free to try it and share results or feedback. 🙂 + +─────────────────────────────────────────────────────────────────────────────── + VERSION NOTES +─────────────────────────────────────────────────────────────────────────────── + KR2 _ V2 is for professionalism, not for candid or amateur quality photos. + +─────────────────────────────────────────────────────────────────────────────── + EXAMPLE MEDIA +─────────────────────────────────────────────────────────────────────────────── + 6 preview(s) in ./example_images/ — community results for + this LoRA. Videos keep a matching _poster.jpg still frame. + + No prompts are recorded below: CivitAI's public API flags these posts as + having generation metadata but does not return it, so the prompts cannot be + scraped. Open the model page to read them. + + 01_image.jpeg (image, 832x1248) + 02_image.jpeg (image, 768x1024) + 03_image.jpeg (image, 768x1376) + 04_image.jpeg (image, 2496x1824) + 05_image.jpeg (image, 2496x1824) + 06_image.jpeg (image, 768x1376) + +─────────────────────────────────────────────────────────────────────────────── + HOW TO USE (COMFYUI) +─────────────────────────────────────────────────────────────────────────────── + 1. Update ComfyUI to at least v0.26.0 (released 2026-06-23), which added + local Krea2 support. For reference-image / style-reference / edit LoRAs you + need at least v0.29.0 (2026-07-29), which added regular and timestep-zero + reference images for Krea 2. ModelMergeKrea2 arrived in v0.27.0 + (2026-06-30). + + 2. Put the base stack in place (files from huggingface.co/Comfy-Org/Krea-2, + which is not gated): + - ComfyUI/models/diffusion_models/krea2_turbo_fp8_scaled.safetensors + (documented recommended default; the repo also holds krea2_turbo_bf16, + _int8_convrot, _mxfp8, _nvfp4 and krea2_raw_bf16 / _fp8_scaled / + _int8_convrot) + - ComfyUI/models/text_encoders/qwen3vl_4b_fp8_scaled.safetensors (bf16 + also available) + - ComfyUI/models/vae/qwen_image_vae.safetensors + + 3. Put the LoRA .safetensors in ComfyUI/models/loras/ - flat, no per-model + subfolder required. + + 4. Build or load the graph. The official image_krea2_turbo_t2i.json + template is: UNETLoader (krea2_turbo_fp8_scaled.safetensors) -> + LoraLoaderModelOnly -> KSampler; CLIPLoader + (qwen3vl_4b_fp8_scaled.safetensors, type "krea2") -> text encode -> + KSampler positive, with ConditioningZeroOut into negative; VAELoader + (qwen_image_vae.safetensors) -> VAEDecode -> SaveImage; EmptyLatentImage + 1024x1024 driven by ResolutionSelector. An int8 sibling template + (image_krea2_turbo_t2i_int8.json) also ships. + + 5. Load the LoRA with LoraLoaderModelOnly, not LoraLoader - both official + templates do. Krea 2 LoRAs patch the DiT only; note that the DiT's own + text_fusion / txtfusion blocks are a normal LoRA target (the official style + LoRAs patch them), but the Qwen3-VL text encoder itself is not. Set + strength 1.0 for Krea's official style LoRAs (the shipped template uses + 0.8) and add the LoRA's trigger word to the prompt. + + 6. Sampler settings: steps 8, cfg 1, sampler euler, scheduler simple, + denoise 1.0. Do not add a ModelSampling node for plain t2i - ComfyUI + already applies shift 1.15 from Krea2.sampling_settings. + + 7. For style reference, use the dedicated template: UNETLoader + krea2_turbo_int8_convrot.safetensors, LoraLoaderModelOnly + krea2_style_reference.safetensors at 1.0, ModelSamplingFlux (1.15, 0.5), + TextEncodeQwenImageEditPlus fed by LoadImage, + FluxKontextMultiReferenceLatentMethod set to index_timestep_zero, and + SamplerCustomAdvanced with KSamplerSelect euler + BasicScheduler simple/8 + + CFGGuider cfg 1. + + 8. For instruction editing (Krea 2 Identity Edit), stock nodes are not + enough: install the ComfyUI-Krea2Edit node pack, which ships its own + workflows, and run Turbo at 8-12 steps / cfg 1 (or Raw at 20 steps / cfg 3 + for removals) with LoRA strength 1.0. + + 9. Low-VRAM alternative: install city96's ComfyUI-GGUF and use a community + GGUF build (realrebelai/KREA-2_GGUFs, quantized from Turbo; + molbal/krea2-gguf, quantized from Raw) with the Unet Loader (GGUF) node in + place of UNETLoader. There is no official GGUF release. LoRA loading is + unchanged. + +─────────────────────────────────────────────────────────────────────────────── + Scraped by lorakit on 2026-09-05 00:12 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/ultrareal-krea2-klein9b/metadata.json b/ultrareal-krea2-klein9b/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..3fb2ab32fc06260cd29310ea0a1e80b5f9a62bec --- /dev/null +++ b/ultrareal-krea2-klein9b/metadata.json @@ -0,0 +1,137 @@ +{ + "name": "UltraReal - Krea2, Klein9b", + "slug": "ultrareal-krea2-klein9b", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 2462105, + "version_id": 3182356, + "version_name": "KR2_V2", + "source": "civitai", + "url": "https://civitai.com/models/2462105", + "repo_id": null, + "file_path": null, + "author": "vizsumit", + "category": { + "key": "photoreal", + "label": "Photorealism" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [], + "recommended_settings": { + "LoRA strength": "0.6–0.8" + }, + "chaining": { + "author": { + "stacks": "not stated", + "notes": "", + "avoid": [], + "requires": [], + "order": "", + "strength": "", + "stated": false + }, + "general_rules": [ + "The most important rule is not about chaining but about which checkpoint you chain onto: the official README says in bold \"TRAIN on Raw and RUN on Turbo\", and \"We highly recommend using RAW for training LoRAs and applying them on Turbo for inference.\" LoRAs trained on Raw are stated to work well on Turbo.", + "Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 LoRAs are DiT-only, so there is no CLIP side to chain and the text-encoder branch stays untouched however many LoRAs you add.", + "Krea 2 LoRAs stack in practice and the community treats multi-LoRA as routine - style + character, style + slider, realism + character. Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one writes \"I recently stacked over 20 of my own LoRAs on a single generation and the output still came out great.\"", + "Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is loaded. The CivitAI 12GB training guide also notes rank affects composability: \"rank 8 produced a lighter style influence that was easier to stack with other LoRAs, rank 16 provided a useful balance between concept strength and composability, and rank 32 was better suited to a precise object or concept intended to take priority over other stacked LoRAs.\"", + "Accelerator LoRAs are position-sensitive in intent, not in node order: a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to-Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is explicitly recommended by the 4-step LoRA's author.", + "Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card says \"Keep mu = 1.15. The training targets are anchored to that grid; a different shift moves the [sampler off them]\"; the same reasoning applies to any Turbo-trained adapter in the chain.", + "Reference/edit LoRAs are designed to be stacked under ordinary subject and style LoRAs - the Identity Edit card states \"Composes with your LoRAs: character/body/style LoRAs stack on top and steer the prior\" and recommends a stacked subject LoRA when unusual subjects drift toward the base prior.", + "Krea 2 LoRAs cannot be chained with LoRAs for any other base model. Many CivitAI pages carry multi-model titles like \"(Krea-2 + ZIT)\" or \"[Flux | ZIB | Krea2]\" - those are separate files for separate architectures shipped under one model page, not a combined adapter.", + "To combine two full Krea 2 checkpoints rather than adapters, ComfyUI ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per-section weights for first., tmlp., txtmlp., tproj., txtfusion.projector, txtfusion.layerwise_blocks.0-1, txtfusion.refiner_blocks.0-1, blocks.0-27 and last.", + "The author never states which weight family this targets. Silent degradation, never a hard error, between Raw and Turbo. The two share one architecture - ai-toolkit's config is commented \"The reference 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)\" - so every LoRA tensor shape matches on both and ComfyUI/diffuser" + ] + }, + "pros": [ + "Widely adopted — 36,247 downloads. [source: CivitAI stats]", + "Positively received — 1,104 thumbs-up with no down-votes. [source: CivitAI stats]", + "Compact at 218 MB — cheap to stack with other LoRAs. [source: file size]", + "Author published concrete recommended settings. [source: description]" + ], + "cons": [ + "Author references NSFW output — review before using in shared work. [source: description]", + "No trigger word — the LoRA is always on once loaded, so strength is the only control. [source: model metadata]", + "Author never states which generation mode it was trained for; test against your own workflow before relying on it. [source: description]" + ], + "file": { + "name": "ultra_real_krea2_v2.safetensors", + "size_bytes": 228588904, + "size_mb": 218.0, + "sha256": "233a8b1df4b3387f9f2bedaa2099d0e14cc946d1105f732de2a8600310b86f07" + }, + "stats": { + "downloads": 36247, + "thumbs_up": 1104, + "thumbs_down": 0 + }, + "popularity": { + "downloads": 36247, + "thumbs_up": 1104, + "thumbs_down": 0, + "comments": 0, + "approval": "100% positive (1104/1104)", + "age_days": 36.2, + "downloads_per_day": 196.54, + "rank_by_downloads": 14, + "rank_by_momentum": 11, + "of_total": 150, + "adoption_tier": "Top tier", + "momentum": "Hot · fresh release" + }, + "selection_score": 49495.0, + "selection_reasons": [ + "36,247 downloads", + "1,104 thumbs-up", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "832x1248", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "768x1024", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "768x1376", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "2496x1824", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "2496x1824", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "768x1376", + "prompt": "" + } + ], + "tags": [ + "style", + "photography", + "realism", + "realistic", + "klein" + ], + "scraped_at": "2026-09-05T00:12:07+00:00", + "path": "krea-2/ultrareal-krea2-klein9b" +} diff --git a/ultrareal-krea2-klein9b/ultra_real_krea2_v2.safetensors b/ultrareal-krea2-klein9b/ultra_real_krea2_v2.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..9d406384a162e1533285a6cd7bdf26ae359c162c --- /dev/null +++ b/ultrareal-krea2-klein9b/ultra_real_krea2_v2.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:233a8b1df4b3387f9f2bedaa2099d0e14cc946d1105f732de2a8600310b86f07 +size 228588904