diff --git a/.gitattributes b/.gitattributes index 0f764e29f5650fe38865cac41b7d9546e9007a68..f54ff440a4ec112c74749fb7b7ba55425f6a6021 100644 --- a/.gitattributes +++ b/.gitattributes @@ -244,3 +244,74 @@ krea-2-detail-slider/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs krea-2-detail-slider/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text krea-2-detail-slider/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text krea-2-detail-slider/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +cutifyier/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +cutifyier/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +cutifyier/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +cutifyier/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +cutifyier/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +cutifyier/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +krea-2-identity-edit/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +krea-2-identity-edit/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +krea-2-identity-edit/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +krea-2-identity-edit/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +krea-2-identity-edit/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +krea-2-identity-edit/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +ars-midjourney-watercolor/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +ars-midjourney-watercolor/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +ars-midjourney-watercolor/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +ars-midjourney-watercolor/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +ars-midjourney-watercolor/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +ars-midjourney-watercolor/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +impressionism/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +impressionism/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +impressionism/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +impressionism/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +impressionism/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +impressionism/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +carefull-style-krea-2-anima-il/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +carefull-style-krea-2-anima-il/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +carefull-style-krea-2-anima-il/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +carefull-style-krea-2-anima-il/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +carefull-style-krea-2-anima-il/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +carefull-style-krea-2-anima-il/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +the-age-slider-zit-krea-2/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +the-age-slider-zit-krea-2/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +the-age-slider-zit-krea-2/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +the-age-slider-zit-krea-2/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +the-age-slider-zit-krea-2/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +the-age-slider-zit-krea-2/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +frank-frazetta-style-oil-painting-krea2-flux/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +frank-frazetta-style-oil-painting-krea2-flux/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +frank-frazetta-style-oil-painting-krea2-flux/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +frank-frazetta-style-oil-painting-krea2-flux/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +frank-frazetta-style-oil-painting-krea2-flux/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +frank-frazetta-style-oil-painting-krea2-flux/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +omar-dogan-style-1-5-pony-illustrious-noob-zit-anima-krea2/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +omar-dogan-style-1-5-pony-illustrious-noob-zit-anima-krea2/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +omar-dogan-style-1-5-pony-illustrious-noob-zit-anima-krea2/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +omar-dogan-style-1-5-pony-illustrious-noob-zit-anima-krea2/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +omar-dogan-style-1-5-pony-illustrious-noob-zit-anima-krea2/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +omar-dogan-style-1-5-pony-illustrious-noob-zit-anima-krea2/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +60s-psychedelic-movie/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +60s-psychedelic-movie/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +60s-psychedelic-movie/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +60s-psychedelic-movie/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +60s-psychedelic-movie/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +60s-psychedelic-movie/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +70s-sci-fi-movie/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +70s-sci-fi-movie/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +70s-sci-fi-movie/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +70s-sci-fi-movie/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +70s-sci-fi-movie/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +70s-sci-fi-movie/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +body-weight-slider-skinny-fat-slider/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +body-weight-slider-skinny-fat-slider/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +body-weight-slider-skinny-fat-slider/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +body-weight-slider-skinny-fat-slider/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +body-weight-slider-skinny-fat-slider/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +body-weight-slider-skinny-fat-slider/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +anna-yamada-and-other-the-dangers-in-my-heart-characters/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +anna-yamada-and-other-the-dangers-in-my-heart-characters/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +anna-yamada-and-other-the-dangers-in-my-heart-characters/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +anna-yamada-and-other-the-dangers-in-my-heart-characters/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +anna-yamada-and-other-the-dangers-in-my-heart-characters/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text diff --git a/60s-psychedelic-movie/60sPsyKrea2Raw.safetensors b/60s-psychedelic-movie/60sPsyKrea2Raw.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..fd484d498a639e60642b7b98e0786696875e8f4d --- /dev/null +++ b/60s-psychedelic-movie/60sPsyKrea2Raw.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd1d53eb64721ef82770c8364d9f27e9f56d56eaa88c263bc5f0781eaba9765b +size 228587728 diff --git a/60s-psychedelic-movie/example_images/01_image.jpeg b/60s-psychedelic-movie/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..b302a8a80e7082b8c7b73290b6ac3869276258e6 --- /dev/null +++ b/60s-psychedelic-movie/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a82b585802a4c3ff35a27cd9da014687fd95b6a742b1993d9ade6235752484ba +size 1710782 diff --git a/60s-psychedelic-movie/example_images/02_image.jpeg b/60s-psychedelic-movie/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..7906d997539915bfa9e6075151a06bc83171e2ac --- /dev/null +++ b/60s-psychedelic-movie/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea4f735e17ebb3cc402fe878a4e19b0913ea2f8d188414045a879437a7d0b26 +size 1689972 diff --git a/60s-psychedelic-movie/example_images/03_image.jpeg b/60s-psychedelic-movie/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..2605066333e15fd33254c320f852b771d14bc2a7 --- /dev/null +++ b/60s-psychedelic-movie/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fe15edded1255416cbb7d2eba529ebfb8864baab217380dad9087d08230d1ddb +size 1724170 diff --git a/60s-psychedelic-movie/example_images/04_image.jpeg b/60s-psychedelic-movie/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..67f643560c626ef14991c3a52a56c65fdb58da81 --- /dev/null +++ b/60s-psychedelic-movie/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:23772d2277c3f98f17285672a64987dcb420b19973b302b35269be122f84f4e9 +size 4595279 diff --git a/60s-psychedelic-movie/example_images/05_image.jpeg b/60s-psychedelic-movie/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..e0cc5d4ef9b11558f4aaea3020fe25ea8e2a4bb3 --- /dev/null +++ b/60s-psychedelic-movie/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:17d6ab907ba2d64285481240001959dd7e513c817c7e218840836518c4ef9879 +size 8492985 diff --git a/60s-psychedelic-movie/example_images/06_image.jpeg b/60s-psychedelic-movie/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..75fed73ed072ad0c0a94041a04234a0797557032 --- /dev/null +++ b/60s-psychedelic-movie/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1013a02787e468f03c32a2eb86f0068cfc1a6e373bdf87d6c5e4ea394d5a82d5 +size 8908153 diff --git a/60s-psychedelic-movie/info.txt b/60s-psychedelic-movie/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..a9811d22f74866ab6b7b5ea4be214d43c1b7d2ee --- /dev/null +++ b/60s-psychedelic-movie/info.txt @@ -0,0 +1,284 @@ +═══════════════════════════════════════════════════════════════════════════════ + 60S PSYCHEDELIC MOVIE +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Visual style + Author ArsMachina + Source https://civitai.com/models/878199 + Provenance CivitAI · model 878199 · version 3129133 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File 60sPsyKrea2Raw.safetensors + Version Krea2 + Size 218.0 MB + Format SafeTensor / unknown precision + SHA256 fd1d53eb64721ef82770c8364d9f27e9f56d56eaa88c263bc5f0781eaba9765b + Published 2026-07-30 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + › "ArsMovieStill, movie still from a 60s psychedelic movie," + +─────────────────────────────────────────────────────────────────────────────── + 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 11,997 + Rating 100% positive (870/870) + Comments 0 + Published 37 days ago + Download rate 25.6/day since release + + Adoption rank #46 of 150 in this collection (Moderate) + Momentum rank #86 of 150 by download rate (Long tail · 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 — 11,997 downloads. [source: CivitAI stats] + + Positively received — 870 thumbs-up with no down-votes. [source: + CivitAI stats] + + Explicit trigger word(s) — ArsMovieStill, movie still from a 60s + psychedelic movie, — 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 tries to take you into a 60s psychedelic movie + + The trigger words are ArsMovieStill, movie still from a 60s psychedelic + movie + MJ7 Update coming soon: Reworking the dataset from scratch. Enhanced period + accuracy. + + V2 Update: The training dataset has been expanded by 300%, significantly + increasing its scope. + This update enhances the range of scenes and characters, offering greater + flexibility. + +─────────────────────────────────────────────────────────────────────────────── + 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, 1024x1024) + 03_image.jpeg (image, 1024x1024) + 04_image.jpeg (image, 1664x2432) + 05_image.jpeg (image, 1728x2160) + 06_image.jpeg (image, 2160x2704) + +─────────────────────────────────────────────────────────────────────────────── + 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 00:13 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/60s-psychedelic-movie/metadata.json b/60s-psychedelic-movie/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..e74f6220598ec96055e1e70adc4a58b9a466397a --- /dev/null +++ b/60s-psychedelic-movie/metadata.json @@ -0,0 +1,134 @@ +{ + "name": "60s Psychedelic Movie", + "slug": "60s-psychedelic-movie", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 878199, + "version_id": 3129133, + "version_name": "Krea2", + "source": "civitai", + "url": "https://civitai.com/models/878199", + "repo_id": null, + "file_path": null, + "author": "ArsMachina", + "category": { + "key": "style", + "label": "Visual style" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [ + "ArsMovieStill, movie still from a 60s psychedelic movie," + ], + "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 — 11,997 downloads. [source: CivitAI stats]", + "Positively received — 870 thumbs-up with no down-votes. [source: CivitAI stats]", + "Explicit trigger word(s) — ArsMovieStill, movie still from a 60s psychedelic movie, — 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": "60sPsyKrea2Raw.safetensors", + "size_bytes": 228587728, + "size_mb": 218.0, + "sha256": "fd1d53eb64721ef82770c8364d9f27e9f56d56eaa88c263bc5f0781eaba9765b" + }, + "stats": { + "downloads": 11997, + "thumbs_up": 870, + "thumbs_down": 0 + }, + "popularity": { + "downloads": 11997, + "thumbs_up": 870, + "thumbs_down": 0, + "comments": 0, + "approval": "100% positive (870/870)", + "age_days": 36.6, + "downloads_per_day": 25.64, + "rank_by_downloads": 46, + "rank_by_momentum": 86, + "of_total": 150, + "adoption_tier": "Moderate", + "momentum": "Long tail · fresh release" + }, + "selection_score": 22837.0, + "selection_reasons": [ + "11,997 downloads", + "870 thumbs-up", + "documented triggers + usage notes", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "1024x1024", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "1024x1024", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "1024x1024", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "1664x2432", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "1728x2160", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "2160x2704", + "prompt": "" + } + ], + "tags": [ + "movie", + "still", + "style" + ], + "scraped_at": "2026-09-05T00:13:30+00:00", + "path": "krea-2/60s-psychedelic-movie" +} diff --git a/70s-sci-fi-movie/70sSCiFiKrea2Raw.safetensors b/70s-sci-fi-movie/70sSCiFiKrea2Raw.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..a83c96f497c453ebdafe87f7c1204b2ef747563f --- /dev/null +++ b/70s-sci-fi-movie/70sSCiFiKrea2Raw.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8d0cfed436f4dd838238779adc9e1da910021885a8ca402f7a4db169cb185c3c +size 228587728 diff --git a/70s-sci-fi-movie/example_images/01_image.jpeg b/70s-sci-fi-movie/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..3fb152ca6a5285b6c4c2f4d7b0d62d124420e066 --- /dev/null +++ b/70s-sci-fi-movie/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:18c709ab9688234f5e4acaa3ac0269aa69deab8249fe41867e020ee3fa43bd81 +size 1312338 diff --git a/70s-sci-fi-movie/example_images/02_image.jpeg b/70s-sci-fi-movie/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..309275ea933c5116030aadc321cb170dd3d2f140 --- /dev/null +++ b/70s-sci-fi-movie/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bafcb702f6cc851ed44a770a0561ef274938d6e04252d31d7853730883155fee +size 1445878 diff --git a/70s-sci-fi-movie/example_images/03_image.jpeg b/70s-sci-fi-movie/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..c3c746c7814c5c596a6e96ba3a1c31af01f937b7 --- /dev/null +++ b/70s-sci-fi-movie/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5c9f225110f7127d7bada0960215c2d3aaf2d052f3c1e4162f87758817369cd0 +size 1946752 diff --git a/70s-sci-fi-movie/example_images/04_image.jpeg b/70s-sci-fi-movie/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..ee6fbe15dd2bacc5f42bdf74dbd8d34fc2dcade2 --- /dev/null +++ b/70s-sci-fi-movie/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d309f207ab672e3f3ba2553772489bef08a07e3a56f8e062d8d437a51caaa69 +size 1495235 diff --git a/70s-sci-fi-movie/example_images/05_image.jpeg b/70s-sci-fi-movie/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..35c0532f306974ceb3bbf8f38de618fa9e27aa90 --- /dev/null +++ b/70s-sci-fi-movie/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:20cc25862665a8938b837ae7be4f4f909f31e9be6c1ee6b60793b28acd955274 +size 1547954 diff --git a/70s-sci-fi-movie/example_images/06_image.jpeg b/70s-sci-fi-movie/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..d8324b1917dc50d36189e5310d9873a6ce713286 --- /dev/null +++ b/70s-sci-fi-movie/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e85808d8a3b25e7bccbd55f6bb1c47bf55681bebcd6b56e165bb376d674c39b7 +size 1390385 diff --git a/70s-sci-fi-movie/info.txt b/70s-sci-fi-movie/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..79c17a11dea105ac25114bd13e355b70d6671c07 --- /dev/null +++ b/70s-sci-fi-movie/info.txt @@ -0,0 +1,279 @@ +═══════════════════════════════════════════════════════════════════════════════ + 70S SCI-FI MOVIE +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Visual style + Author ArsMachina + Source https://civitai.com/models/995002 + Provenance CivitAI · model 995002 · version 3129771 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File 70sSCiFiKrea2Raw.safetensors + Version Krea2 + Size 218.0 MB + Format SafeTensor / unknown precision + SHA256 8d0cfed436f4dd838238779adc9e1da910021885a8ca402f7a4db169cb185c3c + Published 2026-07-30 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + › "ArsMovieStill, movie still from a 1970s technicolor Sci-Fi movie," + +─────────────────────────────────────────────────────────────────────────────── + 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 12,163 + Rating 100% positive (833/833) + Comments 0 + Published 37 days ago + Download rate 23.7/day since release + + Adoption rank #44 of 150 in this collection (Moderate) + Momentum rank #88 of 150 by download rate (Long tail · 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 — 12,163 downloads. [source: CivitAI stats] + + Positively received — 833 thumbs-up with no down-votes. [source: + CivitAI stats] + + Explicit trigger word(s) — ArsMovieStill, movie still from a 1970s + technicolor Sci-Fi movie, — 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 +─────────────────────────────────────────────────────────────────────────────── + MJ7 Update: A new version recreated using MidJourney V7. + It is more "true" to the era/movie style, but to achieve this I have almost + removed my "signiture look and feel" that is present in V1. + This LoRA aesthetic is aimed at the early 70s Sci-fi movies . + The trigger words are ArsMovieStill, movie still from a 1970s technicolor + Sci-Fi movie + +─────────────────────────────────────────────────────────────────────────────── + 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, 1024x1024) + 03_image.jpeg (image, 1080x1445) + 04_image.jpeg (image, 1024x1024) + 05_image.jpeg (image, 1024x1024) + 06_image.jpeg (image, 1024x1024) + +─────────────────────────────────────────────────────────────────────────────── + 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 00:13 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/70s-sci-fi-movie/metadata.json b/70s-sci-fi-movie/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..1b5e2802a66aac38103fd8e33df51b828a129c08 --- /dev/null +++ b/70s-sci-fi-movie/metadata.json @@ -0,0 +1,139 @@ +{ + "name": "70s Sci-Fi Movie", + "slug": "70s-sci-fi-movie", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 995002, + "version_id": 3129771, + "version_name": "Krea2", + "source": "civitai", + "url": "https://civitai.com/models/995002", + "repo_id": null, + "file_path": null, + "author": "ArsMachina", + "category": { + "key": "style", + "label": "Visual style" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [ + "ArsMovieStill, movie still from a 1970s technicolor Sci-Fi movie," + ], + "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 — 12,163 downloads. [source: CivitAI stats]", + "Positively received — 833 thumbs-up with no down-votes. [source: CivitAI stats]", + "Explicit trigger word(s) — ArsMovieStill, movie still from a 1970s technicolor Sci-Fi movie, — 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": "70sSCiFiKrea2Raw.safetensors", + "size_bytes": 228587728, + "size_mb": 218.0, + "sha256": "8d0cfed436f4dd838238779adc9e1da910021885a8ca402f7a4db169cb185c3c" + }, + "stats": { + "downloads": 12163, + "thumbs_up": 833, + "thumbs_down": 0 + }, + "popularity": { + "downloads": 12163, + "thumbs_up": 833, + "thumbs_down": 0, + "comments": 0, + "approval": "100% positive (833/833)", + "age_days": 36.6, + "downloads_per_day": 23.7, + "rank_by_downloads": 44, + "rank_by_momentum": 88, + "of_total": 150, + "adoption_tier": "Moderate", + "momentum": "Long tail · fresh release" + }, + "selection_score": 22559.0, + "selection_reasons": [ + "12,163 downloads", + "833 thumbs-up", + "documented triggers + usage notes", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "1024x1024", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "1024x1024", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "1080x1445", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "1024x1024", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "1024x1024", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "1024x1024", + "prompt": "" + } + ], + "tags": [ + "midjourney", + "70s", + "movie", + "style", + "sci-fi", + "midjourney v6", + "arsmjstyle", + "still" + ], + "scraped_at": "2026-09-05T00:13:32+00:00", + "path": "krea-2/70s-sci-fi-movie" +} diff --git a/anna-yamada-and-other-the-dangers-in-my-heart-characters/Anna_Yamada__The_Dangers_in_My_Heart__-_Krea_2_epoch_10.safetensors b/anna-yamada-and-other-the-dangers-in-my-heart-characters/Anna_Yamada__The_Dangers_in_My_Heart__-_Krea_2_epoch_10.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..e882cfb3f66612e2cbdc170a0528224ccefac740 --- /dev/null +++ b/anna-yamada-and-other-the-dangers-in-my-heart-characters/Anna_Yamada__The_Dangers_in_My_Heart__-_Krea_2_epoch_10.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:72583cbecec8316c8bec130f61b43ea577d6e33360e6623bb9015eb5f0db3d3e +size 228588904 diff --git a/anna-yamada-and-other-the-dangers-in-my-heart-characters/example_images/01_image.jpeg b/anna-yamada-and-other-the-dangers-in-my-heart-characters/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..7ed043384e9e92e8cf9a9e3d59226c6059085923 --- /dev/null +++ b/anna-yamada-and-other-the-dangers-in-my-heart-characters/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eeda7183ba9b8b06bbd47e2d0b692adcbef4e63d13887016b0970d63b8b85c42 +size 3649913 diff --git a/anna-yamada-and-other-the-dangers-in-my-heart-characters/example_images/02_image.jpeg b/anna-yamada-and-other-the-dangers-in-my-heart-characters/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..930b25e5dfe8b35db64f883c21fb9f1224be3b76 --- /dev/null +++ b/anna-yamada-and-other-the-dangers-in-my-heart-characters/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7f3de5c32b085ff24f1d244bac0790acc6cb89982dcbfc0b4e582bdfc92a4955 +size 2036986 diff --git a/anna-yamada-and-other-the-dangers-in-my-heart-characters/example_images/03_image.jpeg b/anna-yamada-and-other-the-dangers-in-my-heart-characters/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..11b94c2a7a3941070dda48f18474d1d3d06b4d26 --- /dev/null +++ b/anna-yamada-and-other-the-dangers-in-my-heart-characters/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:58a84b7e542b6af51dbb7c8988ab4634185d3419232aef9103d309f6ca06c8ce +size 3571402 diff --git a/anna-yamada-and-other-the-dangers-in-my-heart-characters/example_images/04_image.jpeg b/anna-yamada-and-other-the-dangers-in-my-heart-characters/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..f1935824cbe02ca600c61c4eb79b6e7603e1e9bb --- /dev/null +++ b/anna-yamada-and-other-the-dangers-in-my-heart-characters/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:65fbf7bcf87cab5bb07e619cc76b1ed631fba81c51747b5215eda852bc087204 +size 5765522 diff --git a/anna-yamada-and-other-the-dangers-in-my-heart-characters/example_images/05_image.jpeg b/anna-yamada-and-other-the-dangers-in-my-heart-characters/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..9ace60c51753d918876b029409b3850e697d3d5e --- /dev/null +++ b/anna-yamada-and-other-the-dangers-in-my-heart-characters/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b60e669c13b873f3c34d1b0cc6b21006823a30836c52c600207263c5b81309a9 +size 5863421 diff --git a/anna-yamada-and-other-the-dangers-in-my-heart-characters/info.txt b/anna-yamada-and-other-the-dangers-in-my-heart-characters/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..491223afdd0a70f81ea46bbd9bd269b40facd453 --- /dev/null +++ b/anna-yamada-and-other-the-dangers-in-my-heart-characters/info.txt @@ -0,0 +1,350 @@ +═══════════════════════════════════════════════════════════════════════════════ + ANNA YAMADA AND OTHER THE DANGERS IN MY HEART CHARACTERS (ILLUSTRIOUS, KREA + 2, PONY) +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Character / subject + Author MetalchromeX + Source https://civitai.com/models/630682 + Provenance CivitAI · model 630682 · version 3192556 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File Anna_Yamada__The_Dangers_in_My_Heart__-_Krea_2_epoch_10.safetensors + Version Anna Yamada (Krea 2) + Size 218.0 MB + Format SafeTensor / unknown precision + SHA256 72583cbecec8316c8bec130f61b43ea577d6e33360e6623bb9015eb5f0db3d3e + Published 2026-08-02 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + › "&nn&y&m&d&, 1girl, solo, long hair, black hair, brown eyes, mole on + neck, multiple moles on neck, bangs, large breasts" + › "Anna 4 gym outfit" + › "Anna bear hoodie outfit" + › "Anna black turtleneck with skirt outfit" + › "Anna blue butter outfit" + › "Anna blue tanktop outfit" + › "Anna brown school uniform" + › "Anna cooking class outfit" + › "Anna ghost outfit" + › "Anna gym uniform outfit" + › "Anna gym uniform with blue shirt outfit" + › "Anna karate outfit" + › "Anna orange tanktop outfit" + › "Anna pink dress" + › "Anna pink dress outfit" + › "Anna rain hoodie outfit" + › "Anna red dress outfit" + › "Anna red dress turtleneck outfit" + › "Anna scarf outfit" + › "Anna teal tanktop outfit" + › "Anna TV show sailor outfit" + › "Anna twintail with choker outfit" + › "Anna white floral dress outfit" + › "Anna white hoodie with red hood outfit" + › "Anna white tanktop outfit, sports bra, glasses, cleavage, white sports + bra, off shoulder, jacket, black jacket" + › "Anna white turtleneck with plaid dress outfit" + › "Anna school uniform outfit, school uniform, skirt, black socks, socks, + shirt, cardigan, jacket, pleated skirt, white collar shirt, blue skirt" + +─────────────────────────────────────────────────────────────────────────────── + 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 11,881 + Rating 100% positive (781/781) + Comments 0 + Published 33 days ago + Download rate 5.4/day since release + + Adoption rank #48 of 150 in this collection (Moderate) + Momentum rank #147 of 150 by download rate (Long tail · 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 — 11,881 downloads. [source: CivitAI stats] + + Positively received — 781 thumbs-up with no down-votes. [source: + CivitAI stats] + + Explicit trigger word(s) — &nn&y&m&d&, 1girl, solo, long hair, black + hair, brown eyes, mole on neck, multiple moles on neck, bangs, large + breasts, Anna 4 gym outfit, Anna bear hoodie outfit, Anna black + turtleneck with skirt outfit, Anna blue butter outfit, Anna blue + tanktop outfit, Anna brown school uniform, Anna cooking class outfit, + Anna ghost outfit, Anna gym uniform outfit, Anna gym uniform with blue + shirt outfit, Anna karate outfit, Anna orange tanktop outfit, Anna pink + dress, Anna pink dress outfit, Anna rain hoodie outfit, Anna red dress + outfit, Anna red dress turtleneck outfit, Anna scarf outfit, Anna teal + tanktop outfit, Anna TV show sailor outfit, Anna twintail with choker + outfit, Anna white floral dress outfit, Anna white hoodie with red hood + outfit, Anna white tanktop outfit, sports bra, glasses, cleavage, white + sports bra, off shoulder, jacket, black jacket, Anna white turtleneck + with plaid dress outfit, Anna school uniform outfit, school uniform, + skirt, black socks, socks, shirt, cardigan, jacket, pleated skirt, + white collar shirt, blue skirt — 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 +─────────────────────────────────────────────────────────────────────────────── + ➼Like what I do? Consider throwing me a few dollars at https://ko- + fi.com/dandysombraes or commissioning me + https://www.fiverr.com/metalchromex/make-you-a-stable-diffusion-anime- + character-model so I can continue making cool crap for you all to enjoy for + free! + Update log: 8-2-26: I added a Krea 2 LoRA of Anna. + 7-4-26: Happy 4th of July to my fellow Americans! I added a V3 Sanae Yamada + LoRA. V3 is similar to my V2 version, but has tweaked tags, and I removed + some 'eh' images from the dataset. + 1-9-26: Added a V2 Anna Yamada LoRA. V2 has tweaked tags, I removed some + 'eh' images from the dataset, and made other tweaks. V1 should outperform + the original version in most ways. + 11-14-25: Added a Honoka Hara LoRA. + 11-13-25: Added a Nico Kouda LoRA. + 7-24-25: Added a version of Sanae with a lower step count (640 steps). This + might improve onsite generations. + 6-25-25: I added a third version of Anna Yamada with a very low step count + (1,638 between 655 images). I also removed a small handful of 'eh' images. + This should, in theory, help with the on-site trainer and pairing with + other style LoRA. + 6-7-25: Added a Chihiro Kobayashi Illustrious LoRA. + 6-5-25: Added a Yurine Hanzawa Illustrious LoRA. Despite her limited + appearance toward the end of season 2, she turned out pretty well! + 3-6-25: Added an improved 2nd version of my Sanae Yamada model with a + slightly better/expanded dataset and better training settings. + 2-26-25: Added the new Kana Ichikawa Illustrious model and re-posted the + old Pony version of Kana; unpublished the original Pony listing. + 2-16-25: Added a Moeko Sekine Illustrious LoRA. + 2-14-25: Added a Moeko Sekine Pony LoRA and unpublished the old listing; an + Illustrious remake will be coming soon! + 2-13-25: Renamed the page The Dangers in My Heart characters by + Metalchromex. Added the Sanae Yamada Illustrious LoRA and the Anna Yamada + V2 Illustrious LoRA (V2 makes V1 irrelevant) and unpublished the previous + versions. + +─────────────────────────────────────────────────────────────────────────────── + 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, 1200x1920) + 02_image.jpeg (image, 1200x1920) + 03_image.jpeg (image, 1200x1920) + 04_image.jpeg (image, 2160x2864) + 05_image.jpeg (image, 2160x2864) + +─────────────────────────────────────────────────────────────────────────────── + 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 00:13 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/anna-yamada-and-other-the-dangers-in-my-heart-characters/metadata.json b/anna-yamada-and-other-the-dangers-in-my-heart-characters/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..24727120f5a6275756bf519fa26bed3e275b6db1 --- /dev/null +++ b/anna-yamada-and-other-the-dangers-in-my-heart-characters/metadata.json @@ -0,0 +1,156 @@ +{ + "name": "Anna Yamada and other The Dangers in My Heart characters (Illustrious, Krea 2, Pony)", + "slug": "anna-yamada-and-other-the-dangers-in-my-heart-characters", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 630682, + "version_id": 3192556, + "version_name": "Anna Yamada (Krea 2)", + "source": "civitai", + "url": "https://civitai.com/models/630682", + "repo_id": null, + "file_path": null, + "author": "MetalchromeX", + "category": { + "key": "character", + "label": "Character / subject" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [ + "&nn&y&m&d&, 1girl, solo, long hair, black hair, brown eyes, mole on neck, multiple moles on neck, bangs, large breasts", + "Anna 4 gym outfit", + "Anna bear hoodie outfit", + "Anna black turtleneck with skirt outfit", + "Anna blue butter outfit", + "Anna blue tanktop outfit", + "Anna brown school uniform", + "Anna cooking class outfit", + "Anna ghost outfit", + "Anna gym uniform outfit", + "Anna gym uniform with blue shirt outfit", + "Anna karate outfit", + "Anna orange tanktop outfit", + "Anna pink dress", + "Anna pink dress outfit", + "Anna rain hoodie outfit", + "Anna red dress outfit", + "Anna red dress turtleneck outfit", + "Anna scarf outfit", + "Anna teal tanktop outfit", + "Anna TV show sailor outfit", + "Anna twintail with choker outfit", + "Anna white floral dress outfit", + "Anna white hoodie with red hood outfit", + "Anna white tanktop outfit, sports bra, glasses, cleavage, white sports bra, off shoulder, jacket, black jacket", + "Anna white turtleneck with plaid dress outfit", + "Anna school uniform outfit, school uniform, skirt, black socks, socks, shirt, cardigan, jacket, pleated skirt, white collar shirt, blue skirt" + ], + "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 — 11,881 downloads. [source: CivitAI stats]", + "Positively received — 781 thumbs-up with no down-votes. [source: CivitAI stats]", + "Explicit trigger word(s) — &nn&y&m&d&, 1girl, solo, long hair, black hair, brown eyes, mole on neck, multiple moles on neck, bangs, large breasts, Anna 4 gym outfit, Anna bear hoodie outfit, Anna black turtleneck with skirt outfit, Anna blue butter outfit, Anna blue tanktop outfit, Anna brown school uniform, Anna cooking class outfit, Anna ghost outfit, Anna gym uniform outfit, Anna gym uniform with blue shirt outfit, Anna karate outfit, Anna orange tanktop outfit, Anna pink dress, Anna pink dress outfit, Anna rain hoodie outfit, Anna red dress outfit, Anna red dress turtleneck outfit, Anna scarf outfit, Anna teal tanktop outfit, Anna TV show sailor outfit, Anna twintail with choker outfit, Anna white floral dress outfit, Anna white hoodie with red hood outfit, Anna white tanktop outfit, sports bra, glasses, cleavage, white sports bra, off shoulder, jacket, black jacket, Anna white turtleneck with plaid dress outfit, Anna school uniform outfit, school uniform, skirt, black socks, socks, shirt, cardigan, jacket, pleated skirt, white collar shirt, blue skirt — 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": "Anna_Yamada__The_Dangers_in_My_Heart__-_Krea_2_epoch_10.safetensors", + "size_bytes": 228588904, + "size_mb": 218.0, + "sha256": "72583cbecec8316c8bec130f61b43ea577d6e33360e6623bb9015eb5f0db3d3e" + }, + "stats": { + "downloads": 11881, + "thumbs_up": 781, + "thumbs_down": 0 + }, + "popularity": { + "downloads": 11881, + "thumbs_up": 781, + "thumbs_down": 0, + "comments": 0, + "approval": "100% positive (781/781)", + "age_days": 33.1, + "downloads_per_day": 5.37, + "rank_by_downloads": 48, + "rank_by_momentum": 147, + "of_total": 150, + "adoption_tier": "Moderate", + "momentum": "Long tail · fresh release" + }, + "selection_score": 21653.0, + "selection_reasons": [ + "11,881 downloads", + "781 thumbs-up", + "documented triggers + usage notes", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "1200x1920", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "1200x1920", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "1200x1920", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "2160x2864", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "2160x2864", + "prompt": "" + } + ], + "tags": [ + "anna yamada", + "character", + "sanae yamada", + "moeko sekine", + "the dangers in my heart" + ], + "scraped_at": "2026-09-05T00:13:40+00:00", + "path": "krea-2/anna-yamada-and-other-the-dangers-in-my-heart-characters" +} diff --git a/ars-midjourney-watercolor/WaterColorV2Krea2Raw.safetensors b/ars-midjourney-watercolor/WaterColorV2Krea2Raw.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..83dfa4e84ad39cce22f0dda6e857794f905d284c --- /dev/null +++ b/ars-midjourney-watercolor/WaterColorV2Krea2Raw.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cbcc2cc8cf265b18a8ed5af240bebfa1db7b2635bb21d40750d01e815544eccd +size 228587736 diff --git a/ars-midjourney-watercolor/example_images/01_image.jpeg b/ars-midjourney-watercolor/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..464f2963a9e341883a722395fbdc3e9065e67c6d --- /dev/null +++ b/ars-midjourney-watercolor/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:29d768f931fc288b13ac9e61bbaef3cb09e7b62f9342814dfb29778ea46a13bc +size 11160791 diff --git a/ars-midjourney-watercolor/example_images/02_image.jpeg b/ars-midjourney-watercolor/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..63b804f8f567b5c959330b5b2ba3a93596880b50 --- /dev/null +++ b/ars-midjourney-watercolor/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:24c1268f27ee1b4bb6af1c4926ad256e59221d7903665ed96abbaca22c255c69 +size 11337292 diff --git a/ars-midjourney-watercolor/example_images/03_image.jpeg b/ars-midjourney-watercolor/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..9f9809de876418ae2ee7371d6ba6bfc9ae1f9c6f --- /dev/null +++ b/ars-midjourney-watercolor/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b0f74be8c97a354b7685cbe34bd5a4b250a4b448afc1d3790e7b95c9531a5431 +size 11724982 diff --git a/ars-midjourney-watercolor/example_images/04_image.jpeg b/ars-midjourney-watercolor/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..3893e5fae7f61aabf66212155cce50966d3a5469 --- /dev/null +++ b/ars-midjourney-watercolor/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af793518314417d6899d30925a5b05c7f5051dde0166197b51d4a25dc9ccf89b +size 12701025 diff --git a/ars-midjourney-watercolor/example_images/05_image.jpeg b/ars-midjourney-watercolor/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..e7b99c55fd03e14bcc7d3baec937cbacc4dc64d8 --- /dev/null +++ b/ars-midjourney-watercolor/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f6b5d613780eb5a3a4561c61438905d05c797c66c3c90575ca1a2ba8d0ab8b8f +size 13064849 diff --git a/ars-midjourney-watercolor/example_images/06_image.jpeg b/ars-midjourney-watercolor/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..5259bf8dc4267e8e7330a49b95de47f90a9d7ddb --- /dev/null +++ b/ars-midjourney-watercolor/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fa203a784de222489ddeaf9daf459153204b6cd873760039b86a27b00e06a81a +size 11444864 diff --git a/ars-midjourney-watercolor/info.txt b/ars-midjourney-watercolor/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..8c687ff8b7b30e86cbf5ec88963b5ade2da1d65a --- /dev/null +++ b/ars-midjourney-watercolor/info.txt @@ -0,0 +1,299 @@ +═══════════════════════════════════════════════════════════════════════════════ + ARS MIDJOURNEY WATERCOLOR +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Visual style + Author ArsMachina + Source https://civitai.com/models/551903 + Provenance CivitAI · model 551903 · version 3131742 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File WaterColorV2Krea2Raw.safetensors + Version Krea2 + Size 218.0 MB + Format SafeTensor / unknown precision + SHA256 cbcc2cc8cf265b18a8ed5af240bebfa1db7b2635bb21d40750d01e815544eccd + Published 2026-07-30 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + › "ArsMJStyle, Watercolor," + +─────────────────────────────────────────────────────────────────────────────── + 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 13,719 + Rating 100% positive (931/931) + Comments 0 + Published 36 days ago + Download rate 12.1/day since release + + Adoption rank #41 of 150 in this collection (Moderate) + Momentum rank #128 of 150 by download rate (Long tail · 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 — 13,719 downloads. [source: CivitAI stats] + + Positively received — 931 thumbs-up with no down-votes. [source: + CivitAI stats] + + Explicit trigger word(s) — ArsMJStyle, Watercolor, — 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 +─────────────────────────────────────────────────────────────────────────────── + MJ7 VERSION Update: The dataset was rebuilt entirely from scratch using + MidJourney V7. + The main difference from v1 and v2 is a shift towards more natural colors + and textures. + + The Lora is trained on MidJourney Images that use my "Personalised Style" + + Watercolor What is new in v2: + - Significantly increased the training dataset + + - Added support for a wider range of concepts + + - Removed most of the biases that were present in v1 (closed eyes + everywhere for example) + + - Added 3 different intensity of the colors ( intense , deep , soft colors + ) + + - My personal style is changed a little bit , so the images are more "crisp + and clear" + + Works without the trigger words ArsMJStyle, Watercolor + Works great in the 0.2 - 1.5 range. + Start ~1. + As Watercolor is strong existing style and looks really nice as an abstract + painting, you can go pretty high with the strength and still get a nice + result. + I hope you will enjoy this LoRA as much as I did creating it! + +─────────────────────────────────────────────────────────────────────────────── + 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, 2376x3488) + 02_image.jpeg (image, 2376x3488) + 03_image.jpeg (image, 2376x3488) + 04_image.jpeg (image, 2376x3488) + 05_image.jpeg (image, 2376x3488) + 06_image.jpeg (image, 2376x3488) + +─────────────────────────────────────────────────────────────────────────────── + 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 00:13 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/ars-midjourney-watercolor/metadata.json b/ars-midjourney-watercolor/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..5bb25ddc78e1fe1daad17cbaffb38eb02ae72c4d --- /dev/null +++ b/ars-midjourney-watercolor/metadata.json @@ -0,0 +1,136 @@ +{ + "name": "Ars MidJourney Watercolor", + "slug": "ars-midjourney-watercolor", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 551903, + "version_id": 3131742, + "version_name": "Krea2", + "source": "civitai", + "url": "https://civitai.com/models/551903", + "repo_id": null, + "file_path": null, + "author": "ArsMachina", + "category": { + "key": "style", + "label": "Visual style" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [ + "ArsMJStyle, Watercolor," + ], + "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 — 13,719 downloads. [source: CivitAI stats]", + "Positively received — 931 thumbs-up with no down-votes. [source: CivitAI stats]", + "Explicit trigger word(s) — ArsMJStyle, Watercolor, — 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": "WaterColorV2Krea2Raw.safetensors", + "size_bytes": 228587736, + "size_mb": 218.0, + "sha256": "cbcc2cc8cf265b18a8ed5af240bebfa1db7b2635bb21d40750d01e815544eccd" + }, + "stats": { + "downloads": 13719, + "thumbs_up": 931, + "thumbs_down": 0 + }, + "popularity": { + "downloads": 13719, + "thumbs_up": 931, + "thumbs_down": 0, + "comments": 0, + "approval": "100% positive (931/931)", + "age_days": 36.5, + "downloads_per_day": 12.14, + "rank_by_downloads": 41, + "rank_by_momentum": 128, + "of_total": 150, + "adoption_tier": "Moderate", + "momentum": "Long tail · fresh release" + }, + "selection_score": 25291.0, + "selection_reasons": [ + "13,719 downloads", + "931 thumbs-up", + "documented triggers + usage notes", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "2376x3488", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "2376x3488", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "2376x3488", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "2376x3488", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "2376x3488", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "2376x3488", + "prompt": "" + } + ], + "tags": [ + "midjourney", + "arsmjstyle", + "midjourney v6", + "watercolor", + "style" + ], + "scraped_at": "2026-09-05T00:13:13+00:00", + "path": "krea-2/ars-midjourney-watercolor" +} diff --git a/body-weight-slider-skinny-fat-slider/Krea2Weight_v3.safetensors b/body-weight-slider-skinny-fat-slider/Krea2Weight_v3.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..5496efe5cb5cc1867f8af692201372b1cd6c8619 --- /dev/null +++ b/body-weight-slider-skinny-fat-slider/Krea2Weight_v3.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:19d9b17af57918e4168c77dd279c2aa0d87612eb55a0b516f7e7838e25e3a7bc +size 107435704 diff --git a/body-weight-slider-skinny-fat-slider/example_images/01_image.jpeg b/body-weight-slider-skinny-fat-slider/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..02cf324ab585969a9a4346e1bcd7e5d7eb7c9d16 --- /dev/null +++ b/body-weight-slider-skinny-fat-slider/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc91d40f131f4641bb82d8e7bb5f4e6e9663133c12a96cb96f7c78c67dec984d +size 348977 diff --git a/body-weight-slider-skinny-fat-slider/example_images/02_image.jpeg b/body-weight-slider-skinny-fat-slider/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..0b074941e4210ef9eb29e0021cb56b245073ba2a --- /dev/null +++ b/body-weight-slider-skinny-fat-slider/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b869a39df3299b713bb495a5020e402220c908754b157da44e316b61f9ecac9d +size 582077 diff --git a/body-weight-slider-skinny-fat-slider/example_images/03_image.jpeg b/body-weight-slider-skinny-fat-slider/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..2d741959ad8deb707b1361e2953b4b0384b9b7c3 --- /dev/null +++ b/body-weight-slider-skinny-fat-slider/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:feba409c17203e1c6fda0614c665919945134df19d19824e3b3e35302a6ab00c +size 573711 diff --git a/body-weight-slider-skinny-fat-slider/example_images/04_image.jpeg b/body-weight-slider-skinny-fat-slider/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..bd644f67ae41e0d97e1d1a0da8f8c159af11b7ce --- /dev/null +++ b/body-weight-slider-skinny-fat-slider/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:542cef04c9b65fb630f856c1f05cea2758f497140d9ece666ccfe058c13922bf +size 493356 diff --git a/body-weight-slider-skinny-fat-slider/example_images/05_image.jpeg b/body-weight-slider-skinny-fat-slider/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..54921780dc5e0539088cc9ad3d46c95a72fff992 --- /dev/null +++ b/body-weight-slider-skinny-fat-slider/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd30290f93a3855662231b01dac8235251cfdf301af068526f128d12c2995d8a +size 549528 diff --git a/body-weight-slider-skinny-fat-slider/example_images/06_image.jpeg b/body-weight-slider-skinny-fat-slider/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..6cf66a8e3421ce0836cc23073dcd61624b6d6cca --- /dev/null +++ b/body-weight-slider-skinny-fat-slider/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f7280e84c367e4577477993d9e63f81fa1555c6fb11fe8951eb9a973b42115b +size 2625477 diff --git a/body-weight-slider-skinny-fat-slider/info.txt b/body-weight-slider-skinny-fat-slider/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..fb5c39620066e98034df0a16164c33e6baa317a7 --- /dev/null +++ b/body-weight-slider-skinny-fat-slider/info.txt @@ -0,0 +1,288 @@ +═══════════════════════════════════════════════════════════════════════════════ + BODY WEIGHT SLIDER. (SKINNY ===> FAT SLIDER) +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Detail / anatomy fixer + Author Maruhashi + Source https://civitai.com/models/2513759 + Provenance CivitAI · model 2513759 · version 3129180 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File Krea2Weight_v3.safetensors + Version Krea2 + Size 102.5 MB + Format SafeTensor / unknown precision + SHA256 19d9b17af57918e4168c77dd279c2aa0d87612eb55a0b516f7e7838e25e3a7bc + Published 2026-07-26 + +─────────────────────────────────────────────────────────────────────────────── + 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 11,267 + Rating 100% positive (855/856) + Comments 13 + Published 41 days ago + Download rate 85.9/day since release + + Adoption rank #50 of 150 in this collection (Moderate) + Momentum rank #33 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 — 11,267 downloads. [source: CivitAI stats] + + Positively received — 855 thumbs-up, 1 down. [source: CivitAI stats] + + Compact at 102 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 +─────────────────────────────────────────────────────────────────────────────── + Recommended negative prompt: lowres, bad quality, worst quality, bad + anatomy, jpeg artifacts, signature, watermark, sketch + + I also recommend using this upscaler: + https://openmodeldb.info/models/2x-AnimeSharpV4 + + Usable weight: from -3 to 3+, depending on the checkpoint and additional + LORA used + + No trigger word is needed + +─────────────────────────────────────────────────────────────────────────────── + VERSION NOTES +─────────────────────────────────────────────────────────────────────────────── + Works in wide range from -7 to 7, you can try higher weights depending on + checkpoint and additional lora used + +─────────────────────────────────────────────────────────────────────────────── + 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, 1664x2828) + 02_image.jpeg (image, 1440x1920) + 03_image.jpeg (image, 1440x1920) + 04_image.jpeg (image, 1440x1920) + 05_image.jpeg (image, 1440x1920) + 06_image.jpeg (image, 1344x1920) + +─────────────────────────────────────────────────────────────────────────────── + 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 00:13 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/body-weight-slider-skinny-fat-slider/metadata.json b/body-weight-slider-skinny-fat-slider/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..2a7f9f6ab6f47bad5a3cc16ef8774a4f2bc0ef04 --- /dev/null +++ b/body-weight-slider-skinny-fat-slider/metadata.json @@ -0,0 +1,135 @@ +{ + "name": "Body weight slider. (Skinny ===> Fat slider)", + "slug": "body-weight-slider-skinny-fat-slider", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 2513759, + "version_id": 3129180, + "version_name": "Krea2", + "source": "civitai", + "url": "https://civitai.com/models/2513759", + "repo_id": null, + "file_path": null, + "author": "Maruhashi", + "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 — 11,267 downloads. [source: CivitAI stats]", + "Positively received — 855 thumbs-up, 1 down. [source: CivitAI stats]", + "Compact at 102 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": "Krea2Weight_v3.safetensors", + "size_bytes": 107435704, + "size_mb": 102.46, + "sha256": "19d9b17af57918e4168c77dd279c2aa0d87612eb55a0b516f7e7838e25e3a7bc" + }, + "stats": { + "downloads": 11267, + "thumbs_up": 855, + "thumbs_down": 1 + }, + "popularity": { + "downloads": 11267, + "thumbs_up": 855, + "thumbs_down": 1, + "comments": 13, + "approval": "100% positive (855/856)", + "age_days": 40.7, + "downloads_per_day": 85.92, + "rank_by_downloads": 50, + "rank_by_momentum": 33, + "of_total": 150, + "adoption_tier": "Moderate", + "momentum": "Trending · fresh release" + }, + "selection_score": 21527.0, + "selection_reasons": [ + "11,267 downloads", + "855 thumbs-up", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "1664x2828", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "1440x1920", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "1440x1920", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "1440x1920", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "1440x1920", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "1344x1920", + "prompt": "" + } + ], + "tags": [ + "slider", + "concept", + "illustrious", + "fat", + "weight slider", + "body weight slider", + "skinny" + ], + "scraped_at": "2026-09-05T00:13:37+00:00", + "path": "krea-2/body-weight-slider-skinny-fat-slider" +} diff --git a/carefull-style-krea-2-anima-il/carefull_style_krea_v1.safetensors b/carefull-style-krea-2-anima-il/carefull_style_krea_v1.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..afe22dd5fb9d8e723506f46c01b8205a47462f92 --- /dev/null +++ b/carefull-style-krea-2-anima-il/carefull_style_krea_v1.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1d7dd8770f2e0038c73d6d310935493c18b248ceeadf6174f409d00c74d603d7 +size 228588904 diff --git a/carefull-style-krea-2-anima-il/example_images/01_image.jpeg b/carefull-style-krea-2-anima-il/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..ad640a3e11db6f101971d2d501e43ae65d9dbed8 --- /dev/null +++ b/carefull-style-krea-2-anima-il/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:262dfea6d1bb6f0deeee8ba7a1b44a941b2260619620e0861a5d5bd159001f03 +size 2957708 diff --git a/carefull-style-krea-2-anima-il/example_images/02_image.jpeg b/carefull-style-krea-2-anima-il/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..29b18e8bb40301df1f20290290c670cc867c4ac0 --- /dev/null +++ b/carefull-style-krea-2-anima-il/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3d4fe0c48f1b59f84143f4d4bd9f60f94383c2daa015c390800f1a0f740d419b +size 3421800 diff --git a/carefull-style-krea-2-anima-il/example_images/03_image.jpeg b/carefull-style-krea-2-anima-il/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..c7cdd8e843b4f8481ac82ab9fd0f4fcac46abdb0 --- /dev/null +++ b/carefull-style-krea-2-anima-il/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:75cfb4dfc44b91080132ae89f31171425637b48a2663bacfdbec1d2527c76c29 +size 292354 diff --git a/carefull-style-krea-2-anima-il/example_images/04_image.jpeg b/carefull-style-krea-2-anima-il/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..5106c7d178dedce904a2a509f92bbe5bc004942c --- /dev/null +++ b/carefull-style-krea-2-anima-il/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed43e1a137a9e66f45e7ce1eddf14b3e58e7233e476677bf7d39c9e034507edd +size 210244 diff --git a/carefull-style-krea-2-anima-il/example_images/05_image.jpeg b/carefull-style-krea-2-anima-il/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..bb6209bac5f96fb79a28c3073212810ee49e1187 --- /dev/null +++ b/carefull-style-krea-2-anima-il/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8dec45e73d74d849bb4f6e9cc30e996f40a1d51fb027d5f26be75ad03af6a54b +size 258695 diff --git a/carefull-style-krea-2-anima-il/example_images/06_image.jpeg b/carefull-style-krea-2-anima-il/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..82c3cd433694234521264fa86902e6b3764af2c3 --- /dev/null +++ b/carefull-style-krea-2-anima-il/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a08f2792c15140acf36b7a217a1f69c9360e0367abdd50426fcd6270a4a99bf2 +size 212777 diff --git a/carefull-style-krea-2-anima-il/info.txt b/carefull-style-krea-2-anima-il/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..10abe3f995e4a8d9e6dd71a0a7740875a5f6fe3e --- /dev/null +++ b/carefull-style-krea-2-anima-il/info.txt @@ -0,0 +1,296 @@ +═══════════════════════════════════════════════════════════════════════════════ + CAREFULL STYLE [KREA 2 | ANIMA | IL] +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Visual style + Author carefull_time + Source https://civitai.com/models/1484016 + Provenance CivitAI · model 1484016 · version 3178320 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File carefull_style_krea_v1.safetensors + Version Krea 2 v1.0 + Size 218.0 MB + Format SafeTensor / unknown precision + SHA256 1d7dd8770f2e0038c73d6d310935493c18b248ceeadf6174f409d00c74d603d7 + Published 2026-07-29 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + › "Anime carefull_time style." + +─────────────────────────────────────────────────────────────────────────────── + RECOMMENDED SETTINGS +─────────────────────────────────────────────────────────────────────────────── + Taken verbatim from the author's description: + + Sampler steps 8 + Resolution 1280x1856 + CFG / guidance 1 + Sampler euler +scheduler + +─────────────────────────────────────────────────────────────────────────────── + ADOPTION & POPULARITY +─────────────────────────────────────────────────────────────────────────────── + Downloads 8,828 + Rating 100% positive (1316/1316) + Comments 0 + Published 37 days ago + Download rate 19.1/day since release + + Adoption rank #66 of 150 in this collection (Moderate) + Momentum rank #100 of 150 by download rate (Long tail · 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 — 8,828 downloads. [source: CivitAI stats] + + Positively received — 1,316 thumbs-up with no down-votes. [source: + CivitAI stats] + + Explicit trigger word(s) — Anime carefull_time style. — 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 +─────────────────────────────────────────────────────────────────────────────── + - 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 +─────────────────────────────────────────────────────────────────────────────── + Please leave feedback in the model review or write in DM! + + The LoRA was trained on all my images generated using the Pony Diffusion V6 + XL model in 2024 and then my images generated using the Illustrious model + in 2025 and 2026. This is why it may lean toward the lewd side. + Krea 2 v1.0 All images were generated with the following configuration: + Base model: Krea 2 Turbo bf16 + Resolution: 1280x1856 + sampler: euler + scheduler: beta + steps: 8 + cfg:1 + But I tested other configurations, and all worked well! + Anima v1.0 I didn't test it too much, so please leave your feedback! + I'll leave more details here later. + + IL V2.0 The biggest change - it should be more compatible with all + Illustrious models, not only WAI:v11 + Also, I cleaned the training dataset, so I hope there will be fewer + problems with the hands + IMPORTANT! It appears that this version can produce artifacts on the lips + when you use "Uniform" schedule type. As soon as this is the most popular + type in civitai online generator, here are your options: + - Use DPM++ 2M Karras sampler (set it in advanced settings). This works + reliably even on a low number of steps + + - Or use Euler a sampler with a large steps value (>40). This may not work + each time. + + - Or use less lora strength (0.5-0.7). This may not work each time. + + Sorry for this mess, I tried to retrain the model three times, but it looks + like this is a dataset issue. + IL V1.0 It was trained based on WAI-NSFW-illustrious-SDXL:V11.0 model, so + it may not work so well on other models. + +─────────────────────────────────────────────────────────────────────────────── + VERSION NOTES +─────────────────────────────────────────────────────────────────────────────── + All images were generated with the following configuration: + Base model: Krea 2 Turbo bf16 + Resolution: 1280x1856 + sampler: euler + scheduler: beta + steps: 8 + cfg:1 + But I tested other configurations, and all worked well! + +─────────────────────────────────────────────────────────────────────────────── + 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, 1280x1856) + 02_image.jpeg (image, 1280x1856) + 03_image.jpeg (image, 1376x768) + 04_image.jpeg (image, 832x1248) + 05_image.jpeg (image, 1376x768) + 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 00:13 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/carefull-style-krea-2-anima-il/metadata.json b/carefull-style-krea-2-anima-il/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..48681f7fc70703db78f09d30065f16c87d7b0a56 --- /dev/null +++ b/carefull-style-krea-2-anima-il/metadata.json @@ -0,0 +1,143 @@ +{ + "name": "Carefull Style [Krea 2 | Anima | IL]", + "slug": "carefull-style-krea-2-anima-il", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 1484016, + "version_id": 3178320, + "version_name": "Krea 2 v1.0", + "source": "civitai", + "url": "https://civitai.com/models/1484016", + "repo_id": null, + "file_path": null, + "author": "carefull_time", + "category": { + "key": "style", + "label": "Visual style" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [ + "Anime carefull_time style." + ], + "recommended_settings": { + "Sampler steps": "8", + "Resolution": "1280x1856", + "CFG / guidance": "1", + "Sampler": "euler\nscheduler" + }, + "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 — 8,828 downloads. [source: CivitAI stats]", + "Positively received — 1,316 thumbs-up with no down-votes. [source: CivitAI stats]", + "Explicit trigger word(s) — Anime carefull_time style. — 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": [ + "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": "carefull_style_krea_v1.safetensors", + "size_bytes": 228588904, + "size_mb": 218.0, + "sha256": "1d7dd8770f2e0038c73d6d310935493c18b248ceeadf6174f409d00c74d603d7" + }, + "stats": { + "downloads": 8828, + "thumbs_up": 1316, + "thumbs_down": 0 + }, + "popularity": { + "downloads": 8828, + "thumbs_up": 1316, + "thumbs_down": 0, + "comments": 0, + "approval": "100% positive (1316/1316)", + "age_days": 37.4, + "downloads_per_day": 19.07, + "rank_by_downloads": 66, + "rank_by_momentum": 100, + "of_total": 150, + "adoption_tier": "Moderate", + "momentum": "Long tail · fresh release" + }, + "selection_score": 25020.0, + "selection_reasons": [ + "8,828 downloads", + "1,316 thumbs-up", + "documented triggers + usage notes", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "1280x1856", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "1280x1856", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "1376x768", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "832x1248", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "1376x768", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "768x1376", + "prompt": "" + } + ], + "tags": [ + "style", + "neon", + "disney", + "incase", + "gothic" + ], + "scraped_at": "2026-09-05T00:13:14+00:00", + "path": "krea-2/carefull-style-krea-2-anima-il" +} diff --git a/cutifyier/cutifier_krea2.safetensors b/cutifyier/cutifier_krea2.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..85e12552868f8eef523516ccf0db3e912bf44dec --- /dev/null +++ b/cutifyier/cutifier_krea2.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7187237737fc5a505d32883c4049625d3cb56ad829179bbf27be0966c7d3fae4 +size 228587728 diff --git a/cutifyier/example_images/01_image.jpeg b/cutifyier/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..86db3fe50bac6506fe4934a710852ff8b3a1fb3f --- /dev/null +++ b/cutifyier/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4c878a5fe5ef11ffd3362f0c597d8397872c03bcd52a02a7e682a0cbcbe377cb +size 3360917 diff --git a/cutifyier/example_images/02_image.jpeg b/cutifyier/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..bca0a543c9b1327e62e3e747461b8525b08e62a9 --- /dev/null +++ b/cutifyier/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f28c9ef9ae3917f31e0e12a4c5e6d489f95852492343118e16ed7cdc78823973 +size 3988331 diff --git a/cutifyier/example_images/03_image.jpeg b/cutifyier/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..8a24b4bd39455cae11473c20a94863ee04f0bcdb --- /dev/null +++ b/cutifyier/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c522bd3f70b9d93819cc593220b44c028cfd1fc185fb2e0f7c72cd277b0ed0cc +size 3172140 diff --git a/cutifyier/example_images/04_image.jpeg b/cutifyier/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..9dc1ba9cfe684543477b6bc108c68008522a680a --- /dev/null +++ b/cutifyier/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ca48162175838976f2241dd1ada50d2f6698c354058b56c99c71516b473a13e5 +size 3781800 diff --git a/cutifyier/example_images/05_image.jpeg b/cutifyier/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..52d1d8789af299b9429734ef95b5feb0557efa27 --- /dev/null +++ b/cutifyier/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c434531b162fbe722d6d21b66c6911a571f704af73acf6d20d906e84baf82faf +size 3555468 diff --git a/cutifyier/example_images/06_image.jpeg b/cutifyier/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..055fca394549f9e31289a252f54863e33a89f836 --- /dev/null +++ b/cutifyier/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ad11092d64844ffa984b686050e97f1e9c0e2cff74f12159de6d3a030e2146b +size 2800738 diff --git a/cutifyier/info.txt b/cutifyier/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..707712bf2321d2ea5d285d17cb1668a237047089 --- /dev/null +++ b/cutifyier/info.txt @@ -0,0 +1,299 @@ +═══════════════════════════════════════════════════════════════════════════════ + CUTIFYIER +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Character / subject + Author AI_man2025 + Source https://civitai.com/models/2187487 + Provenance CivitAI · model 2187487 · version 3107521 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File cutifier_krea2.safetensors + Version v2 -krea2 + Size 218.0 MB + Format SafeTensor / unknown precision + SHA256 7187237737fc5a505d32883c4049625d3cb56ad829179bbf27be0966c7d3fae4 + Published 2026-07-07 + +─────────────────────────────────────────────────────────────────────────────── + 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 16,599 + Rating 100% positive (768/769) + Comments 0 + Published 59 days ago + Download rate 205.7/day since release + + Adoption rank #34 of 150 in this collection (High) + Momentum rank #10 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 — 16,599 downloads. [source: CivitAI stats] + + Positively received — 768 thumbs-up, 1 down. [source: CivitAI stats] + + Compact at 218 MB — cheap to stack with other LoRAs. [source: file + size] + +─────────────────────────────────────────────────────────────────────────────── + 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 +─────────────────────────────────────────────────────────────────────────────── + Krea2 v2.0: So yeah I trained again, same data set, though this time i + redid the captions with ai tool kit's auto caption. side note training + krea2 taxes the old 3090 a lot, depressing that 24gb of vram is not really + enough these days. Ok the example images, I used a few random prompts and + some of the old prompts too, the workflow embedded is a lora testing work + flow i created, so may not be of use to anyone. I do have a version of this + lora that did better in some cases, but this one did the best over all. For + nsfw you will want to use a nsfw check point. My training data includes a + few nudes, but its more about the facial structure and shape of the woman + than anything "naughty". Use responsibly, enjoy. As always its free... as + all lora's should be! May be the start of me making more....may be a one + off, lets see how its received. Z-image v1.0: First lora in a LONG time, + and first lora on Z image turbo, I followed the video tutorial from the + AItoolkit dev (Very cool tool!) and used my dataset from flux. This makes + women more cute, subjectively in flux it was made to combat flux chin.... + and failed a bit at that, in this its more of a test, but I did notice + womans faces all seemed to generate around late 20's early 30's so this + dataset has more early 20's (and 18, 19) faces. The trigger word is ukgirl + (not sure if its needed or not) Note: I have not had time to do more than a + few generations, though hit no snags in my tests. + +─────────────────────────────────────────────────────────────────────────────── + VERSION NOTES +─────────────────────────────────────────────────────────────────────────────── + updated to Krea2. No triger word needed + +─────────────────────────────────────────────────────────────────────────────── + 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, 1320x1984) + 02_image.jpeg (image, 1320x1984) + 03_image.jpeg (image, 1080x1920) + 04_image.jpeg (image, 1320x1984) + 05_image.jpeg (image, 1240x1864) + 06_image.jpeg (image, 1080x1920) + +─────────────────────────────────────────────────────────────────────────────── + 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 00:13 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/cutifyier/metadata.json b/cutifyier/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..6940bd8f95e2f66f1d4bf90ec58f5fb66d0d8f9d --- /dev/null +++ b/cutifyier/metadata.json @@ -0,0 +1,133 @@ +{ + "name": "Cutifyier", + "slug": "cutifyier", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 2187487, + "version_id": 3107521, + "version_name": "v2 -krea2", + "source": "civitai", + "url": "https://civitai.com/models/2187487", + "repo_id": null, + "file_path": null, + "author": "AI_man2025", + "category": { + "key": "character", + "label": "Character / subject" + }, + "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 — 16,599 downloads. [source: CivitAI stats]", + "Positively received — 768 thumbs-up, 1 down. [source: CivitAI stats]", + "Compact at 218 MB — cheap to stack with other LoRAs. [source: file size]" + ], + "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": "cutifier_krea2.safetensors", + "size_bytes": 228587728, + "size_mb": 218.0, + "sha256": "7187237737fc5a505d32883c4049625d3cb56ad829179bbf27be0966c7d3fae4" + }, + "stats": { + "downloads": 16599, + "thumbs_up": 768, + "thumbs_down": 1 + }, + "popularity": { + "downloads": 16599, + "thumbs_up": 768, + "thumbs_down": 1, + "comments": 0, + "approval": "100% positive (768/769)", + "age_days": 59.4, + "downloads_per_day": 205.65, + "rank_by_downloads": 34, + "rank_by_momentum": 10, + "of_total": 150, + "adoption_tier": "High", + "momentum": "Hot" + }, + "selection_score": 25815.0, + "selection_reasons": [ + "16,599 downloads", + "768 thumbs-up", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "1320x1984", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "1320x1984", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "1080x1920", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "1320x1984", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "1240x1864", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "1080x1920", + "prompt": "" + } + ], + "tags": [ + "cute", + "uk", + "woman", + "concept" + ], + "scraped_at": "2026-09-05T00:13:04+00:00", + "path": "krea-2/cutifyier" +} diff --git a/frank-frazetta-style-oil-painting-krea2-flux/example_images/01_image.jpeg b/frank-frazetta-style-oil-painting-krea2-flux/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..29a5501a39580e22fa9534783f2ad728505a0b52 --- /dev/null +++ b/frank-frazetta-style-oil-painting-krea2-flux/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b37bd5a9e2edcdc128aa1c8553aae065ac09fdd1466f5e984ca387551018cfc +size 3212579 diff --git a/frank-frazetta-style-oil-painting-krea2-flux/example_images/02_image.jpeg b/frank-frazetta-style-oil-painting-krea2-flux/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..5c50646e8cd6335f1ba506522c57ac414482e1ad --- /dev/null +++ b/frank-frazetta-style-oil-painting-krea2-flux/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:77cc63a8d4d8fccd059f856052ef254788c137f218e7e626b95b44c3f1cac844 +size 209492 diff --git a/frank-frazetta-style-oil-painting-krea2-flux/example_images/03_image.jpeg b/frank-frazetta-style-oil-painting-krea2-flux/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..c0210e0f51ce0999eb9fcb464ff277e23e9a51bc --- /dev/null +++ b/frank-frazetta-style-oil-painting-krea2-flux/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3d55828db222cdb10cf37c1a3d25070975f3778eaa9434a09edc7c392c1418ba +size 237051 diff --git a/frank-frazetta-style-oil-painting-krea2-flux/example_images/04_image.jpeg b/frank-frazetta-style-oil-painting-krea2-flux/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..009c1cb42997709a00ded2b9ed49d83d365b76e8 --- /dev/null +++ b/frank-frazetta-style-oil-painting-krea2-flux/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8eed10121ae2115f1374b67b4758436b3696d7a47de69c7dc9d65cff954527ab +size 186222 diff --git a/frank-frazetta-style-oil-painting-krea2-flux/example_images/05_image.jpeg b/frank-frazetta-style-oil-painting-krea2-flux/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..ceda6d03f55302551c7ac3acadad177288eaf4da --- /dev/null +++ b/frank-frazetta-style-oil-painting-krea2-flux/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6dafd90fe3a3f1610d1c2d25e2cf9457e1b7457d3f8ef929814cadadd04fd34a +size 377072 diff --git a/frank-frazetta-style-oil-painting-krea2-flux/example_images/06_image.jpeg b/frank-frazetta-style-oil-painting-krea2-flux/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..c79ea3d5dde61aa8709bf0bfa14624dd436058b5 --- /dev/null +++ b/frank-frazetta-style-oil-painting-krea2-flux/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:73bdcc7e18221954a5e620c59718366d3e66daad994c1bbdeb38db587e4eb5a6 +size 296913 diff --git a/frank-frazetta-style-oil-painting-krea2-flux/info.txt b/frank-frazetta-style-oil-painting-krea2-flux/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..d78233d60f74191d7fcabfa0d62dc26c6d248a58 --- /dev/null +++ b/frank-frazetta-style-oil-painting-krea2-flux/info.txt @@ -0,0 +1,256 @@ +═══════════════════════════════════════════════════════════════════════════════ + FRANK FRAZETTA STYLE OIL PAINTING [KREA2, FLUX] +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Visual style + Author darmox + Source https://civitai.com/models/657789 + Provenance CivitAI · model 657789 · version 3096520 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File krea2_frazetta_v1_000002000.safetensors + Version Krea 2 + Size 109.0 MB + Format SafeTensor / unknown precision + SHA256 cbd9f5e94804617cd6471f5fc971d489fe9bfaa343d552e9e92914f307ed660e + Published 2026-07-04 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + › "frazetta style dark fantasy oil painting" + +─────────────────────────────────────────────────────────────────────────────── + RECOMMENDED SETTINGS +─────────────────────────────────────────────────────────────────────────────── + Taken verbatim from the author's description: + + Sampler steps 30–35 + +─────────────────────────────────────────────────────────────────────────────── + ADOPTION & POPULARITY +─────────────────────────────────────────────────────────────────────────────── + Downloads 9,176 + Rating 100% positive (1165/1166) + Comments 0 + Published 63 days ago + Download rate 16.0/day since release + + Adoption rank #60 of 150 in this collection (Moderate) + Momentum rank #112 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 — 9,176 downloads. [source: CivitAI stats] + + Positively received — 1,165 thumbs-up, 1 down. [source: CivitAI stats] + + Explicit trigger word(s) — frazetta style dark fantasy oil painting — + so the effect can be turned on and off from the prompt. [source: model + metadata] + + Compact at 109 MB — cheap to stack with other LoRAs. [source: file + size] + + Author published concrete recommended settings. [source: description] + +─────────────────────────────────────────────────────────────────────────────── + 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 +─────────────────────────────────────────────────────────────────────────────── + Trained to match the style of oil paintings by the legendary Frank + Frazetta. + + Krea 2 + Good Starting Prompt: Frazetta style dark fantasy oil painting. + .9 is great but anything after .5 usually has good results. + Flux v2.0 + Good Starting Prompt: Frank Frazetta fantasy oil painting + 1.25 usually works for a great starting strength, with 2.0 cfg and 30 - 35 + steps + Flux v1.0 + Good Starting Prompt: fantasy oil painting in the style of fr4z3tt4 + .75 usually works for a great starting strenght + +─────────────────────────────────────────────────────────────────────────────── + 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, 1280x1600) + 02_image.jpeg (image, 928x1152) + 03_image.jpeg (image, 1024x1024) + 04_image.jpeg (image, 864x1080) + 05_image.jpeg (image, 1184x896) + 06_image.jpeg (image, 1184x896) + +─────────────────────────────────────────────────────────────────────────────── + 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 00:13 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/frank-frazetta-style-oil-painting-krea2-flux/krea2_frazetta_v1_000002000.safetensors b/frank-frazetta-style-oil-painting-krea2-flux/krea2_frazetta_v1_000002000.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..34540f0c04f825e91a1dbc6c51520b65d36064e6 --- /dev/null +++ b/frank-frazetta-style-oil-painting-krea2-flux/krea2_frazetta_v1_000002000.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cbd9f5e94804617cd6471f5fc971d489fe9bfaa343d552e9e92914f307ed660e +size 114325360 diff --git a/frank-frazetta-style-oil-painting-krea2-flux/metadata.json b/frank-frazetta-style-oil-painting-krea2-flux/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..9cf2b8885fe079ae1cdf01bde837b52e08f19c1d --- /dev/null +++ b/frank-frazetta-style-oil-painting-krea2-flux/metadata.json @@ -0,0 +1,135 @@ +{ + "name": "Frank Frazetta Style Oil Painting [Krea2, Flux]", + "slug": "frank-frazetta-style-oil-painting-krea2-flux", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 657789, + "version_id": 3096520, + "version_name": "Krea 2", + "source": "civitai", + "url": "https://civitai.com/models/657789", + "repo_id": null, + "file_path": null, + "author": "darmox", + "category": { + "key": "style", + "label": "Visual style" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [ + "frazetta style dark fantasy oil painting" + ], + "recommended_settings": { + "Sampler steps": "30–35" + }, + "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 — 9,176 downloads. [source: CivitAI stats]", + "Positively received — 1,165 thumbs-up, 1 down. [source: CivitAI stats]", + "Explicit trigger word(s) — frazetta style dark fantasy oil painting — so the effect can be turned on and off from the prompt. [source: model metadata]", + "Compact at 109 MB — cheap to stack with other LoRAs. [source: file size]", + "Author published concrete recommended settings. [source: description]" + ], + "cons": [ + "Author never states which generation mode it was trained for; test against your own workflow before relying on it. [source: description]" + ], + "file": { + "name": "krea2_frazetta_v1_000002000.safetensors", + "size_bytes": 114325360, + "size_mb": 109.03, + "sha256": "cbd9f5e94804617cd6471f5fc971d489fe9bfaa343d552e9e92914f307ed660e" + }, + "stats": { + "downloads": 9176, + "thumbs_up": 1165, + "thumbs_down": 1 + }, + "popularity": { + "downloads": 9176, + "thumbs_up": 1165, + "thumbs_down": 1, + "comments": 0, + "approval": "100% positive (1165/1166)", + "age_days": 62.9, + "downloads_per_day": 16.04, + "rank_by_downloads": 60, + "rank_by_momentum": 112, + "of_total": 150, + "adoption_tier": "Moderate", + "momentum": "Long tail" + }, + "selection_score": 23556.0, + "selection_reasons": [ + "9,176 downloads", + "1,165 thumbs-up", + "documented triggers + usage notes", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "1280x1600", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "928x1152", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "1024x1024", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "864x1080", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "1184x896", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "1184x896", + "prompt": "" + } + ], + "tags": [ + "style" + ], + "scraped_at": "2026-09-05T00:13:20+00:00", + "path": "krea-2/frank-frazetta-style-oil-painting-krea2-flux" +} diff --git a/impressionism/ImpressionismKrea2Raw.safetensors b/impressionism/ImpressionismKrea2Raw.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..cf101aaea5bd72a80733c5a6f7e111818a48cef3 --- /dev/null +++ b/impressionism/ImpressionismKrea2Raw.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e3ce9140ae0dca6ccdd898efd53c0c2b2cbac37a8decf3fbfd65832f8cfdb17b +size 228587736 diff --git a/impressionism/example_images/01_image.jpeg b/impressionism/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..8c204b57d60f8e2000ea4527f9bf92e9d5c1710b --- /dev/null +++ b/impressionism/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bd9cd6dd3ead860c9734089be185a51b3a45d0b270023e4d4f8fb7afbf275606 +size 354914 diff --git a/impressionism/example_images/02_image.jpeg b/impressionism/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..76b1da5943ddce243dcfe8a56333fd437b8a9ffb --- /dev/null +++ b/impressionism/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:525230eb11f70753d82cc1be83034c69502e9efecb9721fbc2d826266e143b0e +size 1900976 diff --git a/impressionism/example_images/03_image.jpeg b/impressionism/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..2c2d787c18bdba6093d1f5977f306abd1638948b --- /dev/null +++ b/impressionism/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:45aff73aeab654541859b5e376f9a1ab8b811af0cd7fd4f4c04f2bb29e9b966f +size 1832134 diff --git a/impressionism/example_images/04_image.jpeg b/impressionism/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..459e53067223f04355ec55b7ee79f2441d5f7033 --- /dev/null +++ b/impressionism/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0831141ea958467b48d2136268a7d597d0b10bc2d91478d807d012f9fb83d3b1 +size 1893718 diff --git a/impressionism/example_images/05_image.jpeg b/impressionism/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..4eac3febfd1326778664e01187e4be17e2a2824d --- /dev/null +++ b/impressionism/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a4a53235eafa60841355511c4a914748fd7153418014e34eb76e766b3030db7d +size 1955913 diff --git a/impressionism/example_images/06_image.jpeg b/impressionism/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..299a714629f7685f5cd96044d7595eb79877a04d --- /dev/null +++ b/impressionism/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5c138757ca9851b9d3e2a7f04be06bbf96db7e3f63f0202ba9c0ad5545002f12 +size 2692290 diff --git a/impressionism/info.txt b/impressionism/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..87a5dab3e290846fe0a4af82e2689ca7b127567f --- /dev/null +++ b/impressionism/info.txt @@ -0,0 +1,287 @@ +═══════════════════════════════════════════════════════════════════════════════ + IMPRESSIONISM +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Visual style + Author ArsMachina + Source https://civitai.com/models/545264 + Provenance CivitAI · model 545264 · version 3219131 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File ImpressionismKrea2Raw.safetensors + Version Krea2 + Size 218.0 MB + Format SafeTensor / unknown precision + SHA256 e3ce9140ae0dca6ccdd898efd53c0c2b2cbac37a8decf3fbfd65832f8cfdb17b + Published 2026-09-03 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + › "ArsMJStyle, impressionism," + +─────────────────────────────────────────────────────────────────────────────── + 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 12,409 + Rating 100% positive (1013/1013) + Comments 0 + Published 1 days ago + Download rate 143.5/day since release + + Adoption rank #43 of 150 in this collection (Moderate) + Momentum rank #15 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 — 12,409 downloads. [source: CivitAI stats] + + Positively received — 1,013 thumbs-up with no down-votes. [source: + CivitAI stats] + + Explicit trigger word(s) — ArsMJStyle, impressionism, — 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 +─────────────────────────────────────────────────────────────────────────────── + MJ7 Update: + The MJ7 dataset was recreated using MidJourney V7. It has a better + versatility, but it is not as strong as the previous 2 versions. + You can start from higher LoRA strength/weight when using it. + + The Lora is trained on MidJourney Images that use my "Personalised Style" + + Impressionism V2 is highly flexible, but the effects are not as prominent + as V1. + Works without the trigger words ArsMJStyle, Impressionism + + Will start adding visible effects at 0.4+ Works great in the 0.6 - 1.5 + range, depending on your preference + It depends on the model, but you can go pretty high with this Lora's + weight. + Quality tags and negatives reduce the effect. + +─────────────────────────────────────────────────────────────────────────────── + 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, 1184x896) + 02_image.jpeg (image, 1024x1024) + 03_image.jpeg (image, 1024x1024) + 04_image.jpeg (image, 1024x1024) + 05_image.jpeg (image, 1024x1024) + 06_image.jpeg (image, 1408x2048) + +─────────────────────────────────────────────────────────────────────────────── + 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 00:13 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/impressionism/metadata.json b/impressionism/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..4c2bb0e10539c1cf533fbdaf352a624835e71b52 --- /dev/null +++ b/impressionism/metadata.json @@ -0,0 +1,135 @@ +{ + "name": "Impressionism", + "slug": "impressionism", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 545264, + "version_id": 3219131, + "version_name": "Krea2", + "source": "civitai", + "url": "https://civitai.com/models/545264", + "repo_id": null, + "file_path": null, + "author": "ArsMachina", + "category": { + "key": "style", + "label": "Visual style" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [ + "ArsMJStyle, impressionism," + ], + "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 — 12,409 downloads. [source: CivitAI stats]", + "Positively received — 1,013 thumbs-up with no down-votes. [source: CivitAI stats]", + "Explicit trigger word(s) — ArsMJStyle, impressionism, — 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": "ImpressionismKrea2Raw.safetensors", + "size_bytes": 228587736, + "size_mb": 218.0, + "sha256": "e3ce9140ae0dca6ccdd898efd53c0c2b2cbac37a8decf3fbfd65832f8cfdb17b" + }, + "stats": { + "downloads": 12409, + "thumbs_up": 1013, + "thumbs_down": 0 + }, + "popularity": { + "downloads": 12409, + "thumbs_up": 1013, + "thumbs_down": 0, + "comments": 0, + "approval": "100% positive (1013/1013)", + "age_days": 1.4, + "downloads_per_day": 143.54, + "rank_by_downloads": 43, + "rank_by_momentum": 15, + "of_total": 150, + "adoption_tier": "Moderate", + "momentum": "Hot · fresh release" + }, + "selection_score": 24965.0, + "selection_reasons": [ + "12,409 downloads", + "1,013 thumbs-up", + "documented triggers + usage notes", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "1184x896", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "1024x1024", + "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": "1024x1024", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "1408x2048", + "prompt": "" + } + ], + "tags": [ + "impressionism", + "style", + "midjourney", + "midjourney v6" + ], + "scraped_at": "2026-09-05T00:13:13+00:00", + "path": "krea-2/impressionism" +} diff --git a/krea-2-identity-edit/example_images/01_image.jpeg b/krea-2-identity-edit/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..284bf13d5ff025231d179f765abe889af54dadc0 --- /dev/null +++ b/krea-2-identity-edit/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ffe5345484f86f5b75bfa2ad29e1ea21d073dd8270195becf731bfc395423f88 +size 4394064 diff --git a/krea-2-identity-edit/example_images/02_image.jpeg b/krea-2-identity-edit/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..aace30333a6f1a46deefa6c74b64269d1c4a792a --- /dev/null +++ b/krea-2-identity-edit/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5c24c4b495a9b5ad4b90ee191e929ef47c8cdc01bb320fb02f4a7bb68ccf9d04 +size 2429648 diff --git a/krea-2-identity-edit/example_images/03_image.jpeg b/krea-2-identity-edit/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..c20f7ea5ba50c28428dc1d23c72c13459146ef23 --- /dev/null +++ b/krea-2-identity-edit/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:49aa1701b24198f58a99fb3af7cf858911bf2e59a43bf84c316233af75295769 +size 1955976 diff --git a/krea-2-identity-edit/example_images/04_image.jpeg b/krea-2-identity-edit/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..8535c31fc825bc98c25e04e3f61ca7b54a8337dd --- /dev/null +++ b/krea-2-identity-edit/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:39dd43d4e1d7b669efd8533edb6755933fae8bf24b899c09f3cd26668e863049 +size 1493597 diff --git a/krea-2-identity-edit/example_images/05_image.jpeg b/krea-2-identity-edit/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..a3e963e199f4848ce5287dd33b4ecc6442ef404b --- /dev/null +++ b/krea-2-identity-edit/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b959e273cc67daaf749b1704fb6d95728ab069884879049f6c5845205f0c1f3 +size 1151568 diff --git a/krea-2-identity-edit/example_images/06_image.jpeg b/krea-2-identity-edit/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..d6aeda49dfef6a0d0b8f5fb57ab346a417ae19ca --- /dev/null +++ b/krea-2-identity-edit/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:61e3945a2236a99d5d3d2ec762ad92b73d454ae25f098da0841456bb05210d82 +size 1415183 diff --git a/krea-2-identity-edit/info.txt b/krea-2-identity-edit/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..8ddc94cb350d371247485f04e3f32a60728e6021 --- /dev/null +++ b/krea-2-identity-edit/info.txt @@ -0,0 +1,388 @@ +═══════════════════════════════════════════════════════════════════════════════ + KREA 2 IDENTITY EDIT +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode Instruction / reference editing (community, not official) + Checkpoint Krea 2 RAW (krea/Krea-2-Raw) + Category Detail / anatomy fixer + Author conrad_locke + Source https://civitai.com/models/2761113 + Provenance CivitAI · model 2761113 · version 3139172 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File krea2_identity_edit_v1_2.safetensors + Version v1.2 + Size 1,743.6 MB + Format SafeTensor / unknown precision + SHA256 6adf9a69cc9502d286db7b69964d37da7e9cfe4b05b4d004bc275f087d3fd3cf + Published 2026-07-17 + +─────────────────────────────────────────────────────────────────────────────── + 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 1.0 + Sampler steps 8–12 + CFG / guidance 1.0 + +─────────────────────────────────────────────────────────────────────────────── + ADOPTION & POPULARITY +─────────────────────────────────────────────────────────────────────────────── + Downloads 19,907 + Rating 100% positive (984/987) + Comments 149 + Published 49 days ago + Download rate 320.2/day since release + + Adoption rank #31 of 150 in this collection (High) + Momentum rank #6 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. + • 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 — 19,907 downloads. [source: CivitAI stats] + + Positively received — 984 thumbs-up, 3 down. [source: CivitAI stats] + + Author published concrete recommended settings. [source: description] + +─────────────────────────────────────────────────────────────────────────────── + CONS & CAVEATS +─────────────────────────────────────────────────────────────────────────────── + - Author labels this experimental — results vary shot to shot. [source: + description] + - Has a stated dependency — read the author's description before loading. + [source: description] + - No trigger word — the LoRA is always on once loaded, so strength is the + only control. [source: model metadata] + - Large at 1744 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 Identity Edit #v1.2 — new things + In the model: + - Better likeness + + - Character sheets — use and creation + + - Head/face swap (+ eye/person replace) — thanks to stablellama + + - Outpainting + + - Inpainting + + - Try-on + + - Better person removal + + - Higher fidelity + a real high-res pass + + In the nodes: + - ref_boost — reference-fidelity dial ← don't miss this one, try between 2 + and 6 + + - Fixed blurry images / fit geometry (also removes the old "match the + aspect ratio" requirement) + + **Settings:** Turbo, 8–12 steps (8 = composition, 12 = face detail, ~10 + balanced), CFG 1.0, LoRA strength 1.0. Removals/deletions: Two-ref edits: + **scene = image 1, person = image 2.** + Requires the [ComfyUI-Krea2Edit nodes]( + https://github.com/lbouaraba/comfyui-krea2edit ) — see [CHANGELOG]( + https://github.com/lbouaraba/comfyui-krea2edit/blob/main/CHANGELOG.md ). + Instruction-based, identity-preserving image editing for Krea 2 (12.9B + single-stream MMDiT). Give it an image and a plain-language instruction; it + edits while preserving what you didn't ask to change — including the + person. + An unofficial community fine-tune of Krea 2 Raw . Not an official Krea + product; not affiliated with or endorsed by Krea.ai , Inc. + Requires the ComfyUI-Krea2Edit node pack — the LoRA is trained with dual + conditioning (in-context VAE tokens + image-grounded Qwen3-VL encoding) + that stock nodes don't provide. Two ready-made workflows ship with it. + What it does - Person re-staging with likeness: "create a photo of this + person at a night market" — same face, same outfit down to individual moles + and marks, fully relit to the new scene. New camera angles and poses + included. + + - Local edits: recolor, add/remove/replace objects, attribute and outfit + changes, with near-pixel preservation of the rest of the frame. + + - Replace-with-reference: "replace the woman with a big orangutan" — the + replace verb is trained, locality holds. + + - Full-image restyles: global style with preserved composition. + + - Two-input edits (experimental): scene + person as separate references. + Outfits and placement work well; see limitations for faces. + + - Composes with your LoRAs: character/body/style LoRAs stack on top and + steer the prior — something closed editors structurally can't offer. + + Recommended settings Task type Model Steps CFG Most edits (add, recolor, + restyle, re-stage) Turbo 8 1.0 Removals / large deletions Raw 20 3.0 + - Match the output aspect ratio to the source image. Training pairs are + same-size; AR mismatch degrades preservation (edits may apply to only part + of the frame). + + - Generate at ≤2MP. Above that, source content can bleed or subjects + duplicate (training was 768/1024-class). + + - grounding_px is a real dial (trained range 512–1536): lower values = + stronger edit adherence and more uniform scene changes; higher values = + stronger identity/likeness. 768 is a balanced default; try 1024+ for + people. + + - At CFG > 1, ground the negative too (empty prompt + same image). + + - LoRA strength 1.0. + + Known limitations (honest list) - Likeness is texture-faithful, proportion- + conservative. Moles, skin character, hair, and lighting adapt beautifully; + strongly distinctive facial geometry (unusual nose, eye spacing, face + length) regresses toward typical proportions. People whose identity lives + in texture and structure transfer best; geometry-defined faces read as a + "close relative." + + - Two-person inputs keep outfits distinct but faces drift toward each + other. Workaround that works today: chain single-ref inserts (place person + A, then a second edit pass adding person B from their reference). + + - Removal works but is not yet reliable — always use the Raw/CFG 3 recipe; + expect occasional re-renders instead of deletions. + + - Outfit swaps are hit-or-miss — changing what a person wears sometimes + works cleanly and sometimes doesn't apply; reroll or rephrase. + + - Local edits aren't always perfectly local — add/remove/replace operations + can sometimes alter other parts of the frame or shift the overall color + grade. If preservation matters, compare against the source and reroll. + + - Highly unusual visual content (extravagant hairstyles, extreme body + types) can drift toward the base prior — a subject LoRA stacked on top + fixes this. + + License The LoRA weights are a Derivative Model of Krea 2 and are + distributed under the Krea 2 Community License Agreement (see also NOTICE + ). Key points for users: commercial use is permitted under the license's + revenue threshold (§2.3, currently <$1M/yr — above that, contact Krea for + an enterprise license); deployments must implement reasonable content + moderation (§4.2); AI disclosure obligations apply where required (§4.3). + This repository modifies the Krea Model as permitted by §3; it is not + endorsed by Krea. + Research/portfolio release by a self-funded hobbyist. + Showcase All reference people below are themselves AI-generated — no real + likenesses. Prompts are embedded in each image. + +─────────────────────────────────────────────────────────────────────────────── + VERSION NOTES +─────────────────────────────────────────────────────────────────────────────── + v1.2 — new things + In the model: + - Better likeness + + - Character sheets — use and creation + + - Head/face swap (+ eye/person replace) — thanks to stablellama + + - Outpainting + + - Inpainting + + - Try-on + + - Better person removal + + - Higher fidelity + a real high-res pass + + In the nodes: + - ref_boost — reference-fidelity dial ← don't miss this one, try between 2 + and 6 + + - Fixed blurry images / fit geometry (also removes the old "match the + aspect ratio" requirement) + +─────────────────────────────────────────────────────────────────────────────── + 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, 1664x2432) + 02_image.jpeg (image, 1000x1576) + 03_image.jpeg (image, 900x1200) + 04_image.jpeg (image, 1024x1024) + 05_image.jpeg (image, 840x1256) + 06_image.jpeg (image, 2560x3202) + +─────────────────────────────────────────────────────────────────────────────── + 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 00:13 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/krea-2-identity-edit/krea2_identity_edit_v1_2.safetensors b/krea-2-identity-edit/krea2_identity_edit_v1_2.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..adad3703805b59b635800881abb6c6b7450d74c5 --- /dev/null +++ b/krea-2-identity-edit/krea2_identity_edit_v1_2.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6adf9a69cc9502d286db7b69964d37da7e9cfe4b05b4d004bc275f087d3fd3cf +size 1828256432 diff --git a/krea-2-identity-edit/metadata.json b/krea-2-identity-edit/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..b029ffaec585b8a0d9cbbd3f096083de5e312a12 --- /dev/null +++ b/krea-2-identity-edit/metadata.json @@ -0,0 +1,140 @@ +{ + "name": "Krea 2 Identity Edit", + "slug": "krea-2-identity-edit", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 2761113, + "version_id": 3139172, + "version_name": "v1.2", + "source": "civitai", + "url": "https://civitai.com/models/2761113", + "repo_id": null, + "file_path": null, + "author": "conrad_locke", + "category": { + "key": "detail", + "label": "Detail / anatomy fixer" + }, + "modalities": [ + "edit" + ], + "primary_modality": "edit", + "checkpoint_family": "raw", + "trigger_words": [], + "recommended_settings": { + "LoRA strength": "1.0", + "Sampler steps": "8–12", + "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 — 19,907 downloads. [source: CivitAI stats]", + "Positively received — 984 thumbs-up, 3 down. [source: CivitAI stats]", + "Author published concrete recommended settings. [source: description]" + ], + "cons": [ + "Author labels this experimental — results vary shot to shot. [source: description]", + "Has a stated dependency — read the author's description before loading. [source: description]", + "No trigger word — the LoRA is always on once loaded, so strength is the only control. [source: model metadata]", + "Large at 1744 MB — slower to load and heavier to stack. [source: file size]" + ], + "file": { + "name": "krea2_identity_edit_v1_2.safetensors", + "size_bytes": 1828256432, + "size_mb": 1743.56, + "sha256": "6adf9a69cc9502d286db7b69964d37da7e9cfe4b05b4d004bc275f087d3fd3cf" + }, + "stats": { + "downloads": 19907, + "thumbs_up": 984, + "thumbs_down": 3 + }, + "popularity": { + "downloads": 19907, + "thumbs_up": 984, + "thumbs_down": 3, + "comments": 149, + "approval": "100% positive (984/987)", + "age_days": 49.2, + "downloads_per_day": 320.17, + "rank_by_downloads": 31, + "rank_by_momentum": 6, + "of_total": 150, + "adoption_tier": "High", + "momentum": "Hot" + }, + "selection_score": 31715.0, + "selection_reasons": [ + "19,907 downloads", + "984 thumbs-up", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "1664x2432", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "1000x1576", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "900x1200", + "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": "2560x3202", + "prompt": "" + } + ], + "tags": [ + "editing", + "tool", + "identity", + "img2img" + ], + "scraped_at": "2026-09-05T00:13:04+00:00", + "path": "krea-2/krea-2-identity-edit" +} diff --git a/omar-dogan-style-1-5-pony-illustrious-noob-zit-anima-krea2/OmarDogan_Krea2_byKonan.safetensors b/omar-dogan-style-1-5-pony-illustrious-noob-zit-anima-krea2/OmarDogan_Krea2_byKonan.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..42118ed7d68d81e30d5e2ea98f172e2f8aafb8ec --- /dev/null +++ b/omar-dogan-style-1-5-pony-illustrious-noob-zit-anima-krea2/OmarDogan_Krea2_byKonan.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a92281000a58cdbf23d358b5517e1aa07583142353fc8279354e4ea39714a74d +size 228588920 diff --git a/omar-dogan-style-1-5-pony-illustrious-noob-zit-anima-krea2/example_images/01_image.jpeg b/omar-dogan-style-1-5-pony-illustrious-noob-zit-anima-krea2/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..8020e8cb2a123e0f9323edb5757d849ae1f6248b --- /dev/null +++ b/omar-dogan-style-1-5-pony-illustrious-noob-zit-anima-krea2/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:398678c037f0d19345317b2faf5cc9fc0dacfafbb6afe8b8d67f66be51da0b47 +size 9150443 diff --git a/omar-dogan-style-1-5-pony-illustrious-noob-zit-anima-krea2/example_images/02_image.jpeg b/omar-dogan-style-1-5-pony-illustrious-noob-zit-anima-krea2/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..f972091dfa6005cad671ac905b2725e0dc88785a --- /dev/null +++ b/omar-dogan-style-1-5-pony-illustrious-noob-zit-anima-krea2/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:15377f97fc25c9bbacf1c31916c586ca7042070dc416f892bb9ba22cc9e6964c +size 12391784 diff --git a/omar-dogan-style-1-5-pony-illustrious-noob-zit-anima-krea2/example_images/03_image.jpeg b/omar-dogan-style-1-5-pony-illustrious-noob-zit-anima-krea2/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..a2211275fab1da9bbf1f957d1b06b0eff96bf779 --- /dev/null +++ b/omar-dogan-style-1-5-pony-illustrious-noob-zit-anima-krea2/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cc6c267f05254b96a63d9d81376a86775a25b8dfb219a9478a0f0b971cf6b381 +size 13746509 diff --git a/omar-dogan-style-1-5-pony-illustrious-noob-zit-anima-krea2/example_images/04_image.jpeg b/omar-dogan-style-1-5-pony-illustrious-noob-zit-anima-krea2/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..b2a7f0a333b3f9695b8dd4c7c0521f7eb6b5b6ee --- /dev/null +++ b/omar-dogan-style-1-5-pony-illustrious-noob-zit-anima-krea2/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e7f3527017089507d9bdcbdc98c1f4b8cf33c6ec0e1f4f8136e4d53d59ed2240 +size 6321703 diff --git a/omar-dogan-style-1-5-pony-illustrious-noob-zit-anima-krea2/example_images/05_image.jpeg b/omar-dogan-style-1-5-pony-illustrious-noob-zit-anima-krea2/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..cdbb1b9fb51dc3aa2c083820c165a8f8b597f150 --- /dev/null +++ b/omar-dogan-style-1-5-pony-illustrious-noob-zit-anima-krea2/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a99493996bfcb14177645f8a5bd5134b062eb757446ed169d2ce611f8855aa06 +size 5311098 diff --git a/omar-dogan-style-1-5-pony-illustrious-noob-zit-anima-krea2/example_images/06_image.jpeg b/omar-dogan-style-1-5-pony-illustrious-noob-zit-anima-krea2/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..764770f21dab5724f7268f4f74487a327e643edc --- /dev/null +++ b/omar-dogan-style-1-5-pony-illustrious-noob-zit-anima-krea2/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7086afa08c2ff5ac8f4025975ab3f9212598336dd755e5efe1d0c07319285777 +size 5442585 diff --git a/omar-dogan-style-1-5-pony-illustrious-noob-zit-anima-krea2/info.txt b/omar-dogan-style-1-5-pony-illustrious-noob-zit-anima-krea2/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..8fc84717af954404222af0b9af2c3c6d717b4cef --- /dev/null +++ b/omar-dogan-style-1-5-pony-illustrious-noob-zit-anima-krea2/info.txt @@ -0,0 +1,280 @@ +═══════════════════════════════════════════════════════════════════════════════ + OMAR DOGAN - STYLE 1.5/PONY/ILLUSTRIOUS/NOOB/ZIT/ANIMA/KREA2 LORA +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Visual style + Author Konan + Source https://civitai.com/models/16137 + Provenance CivitAI · model 16137 · version 3207200 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File OmarDogan_Krea2_byKonan.safetensors + Version Krea2 + Size 218.0 MB + Format SafeTensor / unknown precision + SHA256 a92281000a58cdbf23d358b5517e1aa07583142353fc8279354e4ea39714a74d + Published 2026-08-07 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + › "traditional media, illustration of" + +─────────────────────────────────────────────────────────────────────────────── + 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 9,034 + Rating 100% positive (1209/1211) + Comments 0 + Published 28 days ago + Download rate 12.4/day since release + + Adoption rank #63 of 150 in this collection (Moderate) + Momentum rank #127 of 150 by download rate (Long tail · 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 — 9,034 downloads. [source: CivitAI stats] + + Positively received — 1,209 thumbs-up, 2 down. [source: CivitAI stats] + + Explicit trigger word(s) — traditional media, illustration of — 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 +─────────────────────────────────────────────────────────────────────────────── + Best result with weight between : 0.8-1. + If you like this style, please check Omar Dogan 's work. + Reviews are really appreciated, i love to see the community use my work, + that's why I share it. + Got a specific request ? I'm open for commission! If you provide enough + data, OCs are accepted + If you want to get updates on my projects as they go, you can follow me on + X + +─────────────────────────────────────────────────────────────────────────────── + 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, 2160x2704) + 02_image.jpeg (image, 2160x2704) + 03_image.jpeg (image, 2160x2704) + 04_image.jpeg (image, 1776x2368) + 05_image.jpeg (image, 1776x2368) + 06_image.jpeg (image, 1776x2368) + +─────────────────────────────────────────────────────────────────────────────── + 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 00:13 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/omar-dogan-style-1-5-pony-illustrious-noob-zit-anima-krea2/metadata.json b/omar-dogan-style-1-5-pony-illustrious-noob-zit-anima-krea2/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..6c9358c48ca15baf3048185b2fca5fb7375f1981 --- /dev/null +++ b/omar-dogan-style-1-5-pony-illustrious-noob-zit-anima-krea2/metadata.json @@ -0,0 +1,138 @@ +{ + "name": "Omar Dogan - Style 1.5/Pony/Illustrious/Noob/ZIT/Anima/Krea2 LORA", + "slug": "omar-dogan-style-1-5-pony-illustrious-noob-zit-anima-krea2", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 16137, + "version_id": 3207200, + "version_name": "Krea2", + "source": "civitai", + "url": "https://civitai.com/models/16137", + "repo_id": null, + "file_path": null, + "author": "Konan", + "category": { + "key": "style", + "label": "Visual style" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [ + "traditional media, illustration of" + ], + "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 — 9,034 downloads. [source: CivitAI stats]", + "Positively received — 1,209 thumbs-up, 2 down. [source: CivitAI stats]", + "Explicit trigger word(s) — traditional media, illustration of — 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": "OmarDogan_Krea2_byKonan.safetensors", + "size_bytes": 228588920, + "size_mb": 218.0, + "sha256": "a92281000a58cdbf23d358b5517e1aa07583142353fc8279354e4ea39714a74d" + }, + "stats": { + "downloads": 9034, + "thumbs_up": 1209, + "thumbs_down": 2 + }, + "popularity": { + "downloads": 9034, + "thumbs_up": 1209, + "thumbs_down": 2, + "comments": 0, + "approval": "100% positive (1209/1211)", + "age_days": 28.3, + "downloads_per_day": 12.42, + "rank_by_downloads": 63, + "rank_by_momentum": 127, + "of_total": 150, + "adoption_tier": "Moderate", + "momentum": "Long tail · fresh release" + }, + "selection_score": 23942.0, + "selection_reasons": [ + "9,034 downloads", + "1,209 thumbs-up", + "documented triggers + usage notes", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "2160x2704", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "2160x2704", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "2160x2704", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "1776x2368", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "1776x2368", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "1776x2368", + "prompt": "" + } + ], + "tags": [ + "omar dogan", + "portraits", + "anime", + "art style", + "style", + "marker", + "artstyle" + ], + "scraped_at": "2026-09-05T00:13:28+00:00", + "path": "krea-2/omar-dogan-style-1-5-pony-illustrious-noob-zit-anima-krea2" +} diff --git a/the-age-slider-zit-krea-2/age_krea2_loraholic.safetensors b/the-age-slider-zit-krea-2/age_krea2_loraholic.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..f8b683d851822d4243154c710c8a903f157a05f7 --- /dev/null +++ b/the-age-slider-zit-krea-2/age_krea2_loraholic.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:43fb1a7dc734f99df01add68c15a1bddb9bfdbca0a8773afa9b4968c03f60a9c +size 7202976 diff --git a/the-age-slider-zit-krea-2/example_images/01_image.jpeg b/the-age-slider-zit-krea-2/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..97b6ac56fa79142d6e0e5c9b22674bbf98623455 --- /dev/null +++ b/the-age-slider-zit-krea-2/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:62606dd28ac7769266f63e34c877abb203e9d5024c645f5732b9e7c0afa87cb9 +size 6291588 diff --git a/the-age-slider-zit-krea-2/example_images/02_image.jpeg b/the-age-slider-zit-krea-2/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..4b59db3a27ea16ef1eba2c2ca1390fca862d78f7 --- /dev/null +++ b/the-age-slider-zit-krea-2/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d2be44406e539e407334d5aef8d1925d9ba73a7bc9da1461f13af5ae1be3300 +size 4846570 diff --git a/the-age-slider-zit-krea-2/example_images/03_image.jpeg b/the-age-slider-zit-krea-2/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..7c3be19afc3d17a8d793da090c0edec44f214e0c --- /dev/null +++ b/the-age-slider-zit-krea-2/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:460c36757f3e0f10f0759a9f794c5b070f52e4aa28bf4ca93fbc6ad33fd91f26 +size 5657275 diff --git a/the-age-slider-zit-krea-2/example_images/04_image.jpeg b/the-age-slider-zit-krea-2/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..6cbf5b3e61e13c567eba9a0955452ddd6ac370a4 --- /dev/null +++ b/the-age-slider-zit-krea-2/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:120ecca9942f9d4a24f8a562e95f91ec920423b9bc007eadf3c4a8af67fafbb2 +size 3154151 diff --git a/the-age-slider-zit-krea-2/example_images/05_image.jpeg b/the-age-slider-zit-krea-2/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..27d09f85723f0f70d40b771c09b2017ede56a55f --- /dev/null +++ b/the-age-slider-zit-krea-2/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f6cd8f9cc1c6a49c636f50d867c6a6d9929edb03163c4f704d75aa7ae150210e +size 4747931 diff --git a/the-age-slider-zit-krea-2/example_images/06_image.jpeg b/the-age-slider-zit-krea-2/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..d52ba9a1f608096dc641e99d0cea8a26ccc92385 --- /dev/null +++ b/the-age-slider-zit-krea-2/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:87406487838b2b0aa7a8766445693096e27883712a18bf2ef3729f43c5008afa +size 6814242 diff --git a/the-age-slider-zit-krea-2/info.txt b/the-age-slider-zit-krea-2/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..a7bc5a3178222fff031c8a6c4825ee14f427f497 --- /dev/null +++ b/the-age-slider-zit-krea-2/info.txt @@ -0,0 +1,295 @@ +═══════════════════════════════════════════════════════════════════════════════ + THE AGE SLIDER - (ZIT + KREA-2) +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Character / subject + Author Loraholic + Source https://civitai.com/models/2533032 + Provenance CivitAI · model 2533032 · version 3067659 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File age_krea2_loraholic.safetensors + Version Krea-2 + Size 6.9 MB + Format SafeTensor / unknown precision + SHA256 43fb1a7dc734f99df01add68c15a1bddb9bfdbca0a8773afa9b4968c03f60a9c + Published 2026-06-25 + +─────────────────────────────────────────────────────────────────────────────── + 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 16,445 + Rating 100% positive (688/689) + Comments 0 + Published 72 days ago + Download rate 174.9/day since release + + Adoption rank #35 of 150 in this collection (High) + Momentum rank #13 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 — 16,445 downloads. [source: CivitAI stats] + + Positively received — 688 thumbs-up, 1 down. [source: CivitAI stats] + + Compact at 7 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 +─────────────────────────────────────────────────────────────────────────────── + DISCLAIMER: ALL Slider LoRAs are trained from prompts ONLY , not from a + dataset of example images. A slider will ONLY create images that the main + model you use already has in its dataset. It will simply force the model in + a direction, in this case the age direction. So use this responsibly and + treat it as a PROMPT you can set a value on. + Krea-2: Rank1, so under 7MB file size. + Can be combined with other LoRAs without any problem, even character LoRAs. + Recommended range: -3 to 8 ZIT: This is THE age slider LoRA that changes + age while leaving the rest of the image completely untouched. Exposure, + color balance, and overall image character are preserved extremely well. + It is about as close to an ideal slider LoRA as I have been able to make. + There is also no Asia bias, etc. + This LoRA can be combined with other LoRAs without any problem. + Recommended range: -3 to 10 Please compare it against the alternatives + yourself. If you end up liking it, a thumbs up is always appreciated. And + if you honestly think other alternatives perform better, feel free to rate + it accordingly. + All my work is 100% free. The only thing that really matters to me is + people leaving a review , good or bad . I will NEVER remove or block anyone + for leaving a bad review, so don't be afraid to give me criticism. + I'm stating this because when a creator blocks someone for leaving a bad + review, that person can no longer see any of the creator's models, + essentially muting their critics. This is something I'm highly critical of, + and I believe it's a fundamental quality problem on this site. + +─────────────────────────────────────────────────────────────────────────────── + 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, 2704x1800) + 02_image.jpeg (image, 2048x2560) + 03_image.jpeg (image, 2048x2560) + 04_image.jpeg (image, 1088x1920) + 05_image.jpeg (image, 1408x2048) + 06_image.jpeg (image, 2048x2560) + +─────────────────────────────────────────────────────────────────────────────── + 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 00:13 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/the-age-slider-zit-krea-2/metadata.json b/the-age-slider-zit-krea-2/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..686e48ab40078d7d6631d41256da5f2e4d8ef5e2 --- /dev/null +++ b/the-age-slider-zit-krea-2/metadata.json @@ -0,0 +1,131 @@ +{ + "name": "THE age slider - (ZIT + Krea-2)", + "slug": "the-age-slider-zit-krea-2", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 2533032, + "version_id": 3067659, + "version_name": "Krea-2", + "source": "civitai", + "url": "https://civitai.com/models/2533032", + "repo_id": null, + "file_path": null, + "author": "Loraholic", + "category": { + "key": "character", + "label": "Character / subject" + }, + "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 — 16,445 downloads. [source: CivitAI stats]", + "Positively received — 688 thumbs-up, 1 down. [source: CivitAI stats]", + "Compact at 7 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": "age_krea2_loraholic.safetensors", + "size_bytes": 7202976, + "size_mb": 6.87, + "sha256": "43fb1a7dc734f99df01add68c15a1bddb9bfdbca0a8773afa9b4968c03f60a9c" + }, + "stats": { + "downloads": 16445, + "thumbs_up": 688, + "thumbs_down": 1 + }, + "popularity": { + "downloads": 16445, + "thumbs_up": 688, + "thumbs_down": 1, + "comments": 0, + "approval": "100% positive (688/689)", + "age_days": 71.8, + "downloads_per_day": 174.9, + "rank_by_downloads": 35, + "rank_by_momentum": 13, + "of_total": 150, + "adoption_tier": "High", + "momentum": "Hot" + }, + "selection_score": 24701.0, + "selection_reasons": [ + "16,445 downloads", + "688 thumbs-up", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "2704x1800", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "2048x2560", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "2048x2560", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "1088x1920", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "1408x2048", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "2048x2560", + "prompt": "" + } + ], + "tags": [ + "concept", + "age", + "slider" + ], + "scraped_at": "2026-09-05T00:13:16+00:00", + "path": "krea-2/the-age-slider-zit-krea-2" +}