diff --git a/.gitattributes b/.gitattributes index 83046773cd5e087ef2bb0467ac14702eb813937e..49082c88b0b09964f83ea022b5b51686dd9b3cee 100644 --- a/.gitattributes +++ b/.gitattributes @@ -729,3 +729,75 @@ chibi-animal-characters/example_images/03_image.jpeg filter=lfs diff=lfs merge=l chibi-animal-characters/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text chibi-animal-characters/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text chibi-animal-characters/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +canon-ultrareal/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +canon-ultrareal/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +canon-ultrareal/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +canon-ultrareal/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +canon-ultrareal/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +canon-ultrareal/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +mikkoani-style/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +mikkoani-style/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +mikkoani-style/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +mikkoani-style/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +mikkoani-style/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +mikkoani-style/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +cyaniji-niji-style-anima-krea2-nai-zit/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +cyaniji-niji-style-anima-krea2-nai-zit/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +cyaniji-niji-style-anima-krea2-nai-zit/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +cyaniji-niji-style-anima-krea2-nai-zit/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +cyaniji-niji-style-anima-krea2-nai-zit/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +cyaniji-niji-style-anima-krea2-nai-zit/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +dark-fantasy-pulp-pinup/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +dark-fantasy-pulp-pinup/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +dark-fantasy-pulp-pinup/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +dark-fantasy-pulp-pinup/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +dark-fantasy-pulp-pinup/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +dark-fantasy-pulp-pinup/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +krea-2-weight-slider/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +krea-2-weight-slider/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +krea-2-weight-slider/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +krea-2-weight-slider/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +krea-2-weight-slider/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +krea-2-weight-slider/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +furry-realistic-style/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +furry-realistic-style/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +furry-realistic-style/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +furry-realistic-style/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +furry-realistic-style/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +furry-realistic-style/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b/example_images/04_video.mp4 filter=lfs diff=lfs merge=lfs -text +rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +748cm-style-krea-2-anima/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +748cm-style-krea-2-anima/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +748cm-style-krea-2-anima/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +748cm-style-krea-2-anima/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +748cm-style-krea-2-anima/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +748cm-style-krea-2-anima/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +meesh-style-krea-2-anima-illustrious/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +meesh-style-krea-2-anima-illustrious/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +meesh-style-krea-2-anima-illustrious/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +meesh-style-krea-2-anima-illustrious/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +meesh-style-krea-2-anima-illustrious/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +meesh-style-krea-2-anima-illustrious/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +bss-styles-for-krea2/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +bss-styles-for-krea2/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +bss-styles-for-krea2/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +bss-styles-for-krea2/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +bss-styles-for-krea2/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +bss-styles-for-krea2/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +delken-furry-realism-enhancer-krea-2-anima-illustrious/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +delken-furry-realism-enhancer-krea-2-anima-illustrious/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +delken-furry-realism-enhancer-krea-2-anima-illustrious/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +delken-furry-realism-enhancer-krea-2-anima-illustrious/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +delken-furry-realism-enhancer-krea-2-anima-illustrious/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +delken-furry-realism-enhancer-krea-2-anima-illustrious/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +hair-length-slider-krea2-zit/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +hair-length-slider-krea2-zit/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +hair-length-slider-krea2-zit/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +hair-length-slider-krea2-zit/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +hair-length-slider-krea2-zit/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +hair-length-slider-krea2-zit/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text diff --git a/748cm-style-krea-2-anima/748_c1-st2000.safetensors b/748cm-style-krea-2-anima/748_c1-st2000.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..f912770c4ef17c7f7dd2928d53c38428487b87ff --- /dev/null +++ b/748cm-style-krea-2-anima/748_c1-st2000.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8fccbb4f022d2f21eb72a276ca8b0f6fd33383fa8ec123e48260e3b4a53fc021 +size 117374840 diff --git a/748cm-style-krea-2-anima/example_images/01_image.jpeg b/748cm-style-krea-2-anima/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..dc2ee19293e1eca5589dcf873794a866e0ea98b0 --- /dev/null +++ b/748cm-style-krea-2-anima/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ef2aff7a6295e22b7156a22d6ae02752a734dab4752926af1f4e865d01b858eb +size 352740 diff --git a/748cm-style-krea-2-anima/example_images/02_image.jpeg b/748cm-style-krea-2-anima/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..8cbb77ae0f30f5cd409ac7fc2ca86c845f791dd4 --- /dev/null +++ b/748cm-style-krea-2-anima/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:211c88c6c3a6f55bfbbf5e83f7b7753b65398bbb1259c3aed47bb78fad1765ad +size 292237 diff --git a/748cm-style-krea-2-anima/example_images/03_image.jpeg b/748cm-style-krea-2-anima/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..90834bcf0cdbf6382fcbfbce461a30184b049811 --- /dev/null +++ b/748cm-style-krea-2-anima/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dcff5d5b7a43c71534ad10318f24f5c7bd5f07d5925cacf5fa00635c93acd9ea +size 13863975 diff --git a/748cm-style-krea-2-anima/example_images/04_image.jpeg b/748cm-style-krea-2-anima/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..1f1c64fb39058d4411bc7ae778af2f34c2212b16 --- /dev/null +++ b/748cm-style-krea-2-anima/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:21a46764933e8cc820fe3ea2cdfef5712341dd82ee41b6b74d79b741ef72c735 +size 3901176 diff --git a/748cm-style-krea-2-anima/example_images/05_image.jpeg b/748cm-style-krea-2-anima/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..bd1001ab0b21b929d135010994dfea9f56835b78 --- /dev/null +++ b/748cm-style-krea-2-anima/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a803e268fe25d95ad43d7668b956d2a89a252506b6a10301e3c5e106a933f84f +size 295764 diff --git a/748cm-style-krea-2-anima/example_images/06_image.jpeg b/748cm-style-krea-2-anima/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..355cbca18f0762518278ca27161cc416cae244ba --- /dev/null +++ b/748cm-style-krea-2-anima/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab90719e3ca1cc755bb079554ed397a9a7cd999a98f5d3e50c513441486f5fde +size 1730885 diff --git a/748cm-style-krea-2-anima/info.txt b/748cm-style-krea-2-anima/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..fe75d3f2f6a6d551349f1f4962b0f2e710dec618 --- /dev/null +++ b/748cm-style-krea-2-anima/info.txt @@ -0,0 +1,251 @@ +═══════════════════════════════════════════════════════════════════════════════ + 748CM STYLE | KREA 2, ANIMA +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Visual style + Author patrik + Source https://civitai.com/models/2626310 + Provenance CivitAI · model 2626310 · version 3180374 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File 748_c1-st2000.safetensors + Version KREA 2 + Size 111.9 MB + Format SafeTensor / unknown precision + SHA256 8fccbb4f022d2f21eb72a276ca8b0f6fd33383fa8ec123e48260e3b4a53fc021 + Published 2026-07-30 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + No trigger word. The LoRA applies as soon as it is loaded — + strength is the only way to dial the effect up or down. + +─────────────────────────────────────────────────────────────────────────────── + RECOMMENDED SETTINGS +─────────────────────────────────────────────────────────────────────────────── + Taken verbatim from the author's description: + + LoRA strength 0.8 + +─────────────────────────────────────────────────────────────────────────────── + ADOPTION & POPULARITY +─────────────────────────────────────────────────────────────────────────────── + Downloads 4,351 + Rating 100% positive (376/376) + Comments 0 + Published 37 days ago + Download rate 21.7/day since release + + Adoption rank #128 of 150 in this collection (Niche) + Momentum rank #94 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 +─────────────────────────────────────────────────────────────────────────────── + + Well established — 4,351 downloads. [source: CivitAI stats] + + Positively received — 376 thumbs-up with no down-votes. [source: + CivitAI stats] + + Compact at 112 MB — cheap to stack with other LoRAs. [source: file + size] + + Author published concrete recommended settings. [source: description] + +─────────────────────────────────────────────────────────────────────────────── + CONS & CAVEATS +─────────────────────────────────────────────────────────────────────────────── + - No trigger word — the LoRA is always on once loaded, so strength is the + only control. [source: model metadata] + - Author never states which generation mode it was trained for; test + against your own workflow before relying on it. [source: description] + +─────────────────────────────────────────────────────────────────────────────── + COMPATIBILITY WARNING +─────────────────────────────────────────────────────────────────────────────── + Silent degradation, never a hard error, between Raw and Turbo. The two + share one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" - so + every LoRA tensor shape matches on both and ComfyUI/diffusers will load an + adapter against either without complaint. The lvladikov 4-step LoRA card + states this for the reverse direction: "Every layer it targets also exists + in Krea 2 Raw, so it will load there without complaint - but that is a side + effect of the shared architecture, not a supported mode", and reports + mixed, subject-dependent results with malformed subjects (duplicated heads, + fused limbs) at low step counts, needing 14+ steps with Raw's CFG on before + output is coherent. The intended direction (train on Raw, run on Turbo) is + officially blessed. The more common practical failure is a settings + mismatch rather than a weight mismatch: running a Turbo-targeted LoRA at + Raw's 52 steps + CFG 3.5, or leaving CFG enabled on Turbo, degrades output + with no warning. Loading a Krea 2 LoRA against a different model family + entirely (Flux, Qwen-Image, Z-Image) is also not an error - ComfyUI logs + "lora key not loaded: {key}" per key and generates as if no LoRA were + applied. + +─────────────────────────────────────────────────────────────────────────────── + AUTHOR'S DESCRIPTION +─────────────────────────────────────────────────────────────────────────────── + 748cm Style LoRA for Anima + Add the distinct "748cm" artistic style to your generations. This LoRA is + trained on preview3 and works best with natural language prompts. + Usage: + - Trigger Word: @748cm_style + + - Recommended Strength: 0.8 (Range: 0.6 - 1.0) + +─────────────────────────────────────────────────────────────────────────────── + EXAMPLE MEDIA +─────────────────────────────────────────────────────────────────────────────── + 6 preview(s) in ./example_images/ — community results for + this LoRA. Videos keep a matching _poster.jpg still frame. + + No prompts are recorded below: CivitAI's public API flags these posts as + having generation metadata but does not return it, so the prompts cannot be + scraped. Open the model page to read them. + + 01_image.jpeg (image, 832x1248) + 02_image.jpeg (image, 928x1152) + 03_image.jpeg (image, 2176x3872) + 04_image.jpeg (image, 1328x1776) + 05_image.jpeg (image, 832x1248) + 06_image.jpeg (image, 848x1520) + +─────────────────────────────────────────────────────────────────────────────── + 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 00:18 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/748cm-style-krea-2-anima/metadata.json b/748cm-style-krea-2-anima/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..74d6706fb4c15dc7679606243f66da8b3d8833aa --- /dev/null +++ b/748cm-style-krea-2-anima/metadata.json @@ -0,0 +1,135 @@ +{ + "name": "748cm Style | KREA 2, Anima", + "slug": "748cm-style-krea-2-anima", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 2626310, + "version_id": 3180374, + "version_name": "KREA 2", + "source": "civitai", + "url": "https://civitai.com/models/2626310", + "repo_id": null, + "file_path": null, + "author": "patrik", + "category": { + "key": "style", + "label": "Visual style" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [], + "recommended_settings": { + "LoRA strength": "0.8" + }, + "chaining": { + "author": { + "stacks": "not stated", + "notes": "", + "avoid": [], + "requires": [], + "order": "", + "strength": "", + "stated": false + }, + "general_rules": [ + "The most important rule is not about chaining but about which checkpoint you chain onto: the official README says in bold \"TRAIN on Raw and RUN on Turbo\", and \"We highly recommend using RAW for training LoRAs and applying them on Turbo for inference.\" LoRAs trained on Raw are stated to work well on Turbo.", + "Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 LoRAs are DiT-only, so there is no CLIP side to chain and the text-encoder branch stays untouched however many LoRAs you add.", + "Krea 2 LoRAs stack in practice and the community treats multi-LoRA as routine - style + character, style + slider, realism + character. Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one writes \"I recently stacked over 20 of my own LoRAs on a single generation and the output still came out great.\"", + "Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is loaded. The CivitAI 12GB training guide also notes rank affects composability: \"rank 8 produced a lighter style influence that was easier to stack with other LoRAs, rank 16 provided a useful balance between concept strength and composability, and rank 32 was better suited to a precise object or concept intended to take priority over other stacked LoRAs.\"", + "Accelerator LoRAs are position-sensitive in intent, not in node order: a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to-Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is explicitly recommended by the 4-step LoRA's author.", + "Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card says \"Keep mu = 1.15. The training targets are anchored to that grid; a different shift moves the [sampler off them]\"; the same reasoning applies to any Turbo-trained adapter in the chain.", + "Reference/edit LoRAs are designed to be stacked under ordinary subject and style LoRAs - the Identity Edit card states \"Composes with your LoRAs: character/body/style LoRAs stack on top and steer the prior\" and recommends a stacked subject LoRA when unusual subjects drift toward the base prior.", + "Krea 2 LoRAs cannot be chained with LoRAs for any other base model. Many CivitAI pages carry multi-model titles like \"(Krea-2 + ZIT)\" or \"[Flux | ZIB | Krea2]\" - those are separate files for separate architectures shipped under one model page, not a combined adapter.", + "To combine two full Krea 2 checkpoints rather than adapters, ComfyUI ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per-section weights for first., tmlp., txtmlp., tproj., txtfusion.projector, txtfusion.layerwise_blocks.0-1, txtfusion.refiner_blocks.0-1, blocks.0-27 and last.", + "The author never states which weight family this targets. Silent degradation, never a hard error, between Raw and Turbo. The two share one architecture - ai-toolkit's config is commented \"The reference 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)\" - so every LoRA tensor shape matches on both and ComfyUI/diffuser" + ] + }, + "pros": [ + "Well established — 4,351 downloads. [source: CivitAI stats]", + "Positively received — 376 thumbs-up with no down-votes. [source: CivitAI stats]", + "Compact at 112 MB — cheap to stack with other LoRAs. [source: file size]", + "Author published concrete recommended settings. [source: description]" + ], + "cons": [ + "No trigger word — the LoRA is always on once loaded, so strength is the only control. [source: model metadata]", + "Author never states which generation mode it was trained for; test against your own workflow before relying on it. [source: description]" + ], + "file": { + "name": "748_c1-st2000.safetensors", + "size_bytes": 117374840, + "size_mb": 111.94, + "sha256": "8fccbb4f022d2f21eb72a276ca8b0f6fd33383fa8ec123e48260e3b4a53fc021" + }, + "stats": { + "downloads": 4351, + "thumbs_up": 376, + "thumbs_down": 0 + }, + "popularity": { + "downloads": 4351, + "thumbs_up": 376, + "thumbs_down": 0, + "comments": 0, + "approval": "100% positive (376/376)", + "age_days": 36.8, + "downloads_per_day": 21.72, + "rank_by_downloads": 128, + "rank_by_momentum": 94, + "of_total": 150, + "adoption_tier": "Niche", + "momentum": "Long tail · fresh release" + }, + "selection_score": 8863.0, + "selection_reasons": [ + "4,351 downloads", + "376 thumbs-up", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "832x1248", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "928x1152", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "2176x3872", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "1328x1776", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "832x1248", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "848x1520", + "prompt": "" + } + ], + "tags": [ + "anime", + "digital illustration", + "748cm", + "style" + ], + "scraped_at": "2026-09-05T00:18:12+00:00", + "path": "krea-2/748cm-style-krea-2-anima" +} diff --git a/bss-styles-for-krea2/DD_krea2_3265864_epoch_4.safetensors b/bss-styles-for-krea2/DD_krea2_3265864_epoch_4.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..b8229a3ba5a932e0e596027650c0fc6e84c17563 --- /dev/null +++ b/bss-styles-for-krea2/DD_krea2_3265864_epoch_4.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ac64c446376ed317dc4c2215c1b8d6e1790b860b78d4361176c29da624ee7080 +size 228588904 diff --git a/bss-styles-for-krea2/example_images/01_image.jpeg b/bss-styles-for-krea2/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..8bb65129187ef9746a4e48f186ca32fdae4c8da6 --- /dev/null +++ b/bss-styles-for-krea2/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:09722905041b8ea550dd0c9ba5e556dd0db4b3f6640e8897526591c1e9edf797 +size 1518174 diff --git a/bss-styles-for-krea2/example_images/02_image.jpeg b/bss-styles-for-krea2/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..78bd8500c496d349c8f8ae0346e3a1b183211a24 --- /dev/null +++ b/bss-styles-for-krea2/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2c885c8d36f53a66dbadd49fdebbaf53ef090ce8b12d3fecc6f3c23103d8309c +size 1218817 diff --git a/bss-styles-for-krea2/example_images/03_image.jpeg b/bss-styles-for-krea2/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..0f52a7357fb0def45511e84da72051b9bff059fc --- /dev/null +++ b/bss-styles-for-krea2/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f8f4716c20c73feea0695a01209f37a00a046eb36d3ff8e9d3663b2ec20b0120 +size 1360270 diff --git a/bss-styles-for-krea2/example_images/04_image.jpeg b/bss-styles-for-krea2/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..09b855523375927df3119cb7aff729b1532c4a1d --- /dev/null +++ b/bss-styles-for-krea2/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:801c22140e6ed9d019862de5a4a6edf507ed2e412594866a6e7127db8388f023 +size 2343753 diff --git a/bss-styles-for-krea2/example_images/05_image.jpeg b/bss-styles-for-krea2/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..30e65afe4893c9b91b6c3ea59a958682bd89a54d --- /dev/null +++ b/bss-styles-for-krea2/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ac3309893bc8a78ec88cc8c583b688f6b14971134c4b4857725846c9298b9f0e +size 1235861 diff --git a/bss-styles-for-krea2/example_images/06_image.jpeg b/bss-styles-for-krea2/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..141cc23f77c38ae8ecdf546ed1fbb44511cef736 --- /dev/null +++ b/bss-styles-for-krea2/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4cb65782100e35fcc6b4360ee2e66d35a8c766a294dce637821cde3acdb60dbc +size 2231127 diff --git a/bss-styles-for-krea2/info.txt b/bss-styles-for-krea2/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..8d019cab3247da8033d78d56f8b7d9db4a352532 --- /dev/null +++ b/bss-styles-for-krea2/info.txt @@ -0,0 +1,298 @@ +═══════════════════════════════════════════════════════════════════════════════ + [BSS] - STYLES FOR KREA2 +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Visual style + Author llikswonskcalb + Source https://civitai.com/models/2762978 + Provenance CivitAI · model 2762978 · version 3268131 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File DD_krea2_3265864_epoch_4.safetensors + Version Digital illustration + Size 218.0 MB + Format SafeTensor / unknown precision + SHA256 ac64c446376ed317dc4c2215c1b8d6e1790b860b78d4361176c29da624ee7080 + Published 2026-08-26 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + › "Digital illustration of" + › "Digital artwork of" + › "Fantasy" + › "Cartoon" + +─────────────────────────────────────────────────────────────────────────────── + 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 4,912 + Rating 100% positive (290/290) + Comments 0 + Published 10 days ago + Download rate 35.8/day since release + + Adoption rank #118 of 150 in this collection (Niche) + Momentum rank #65 of 150 by download rate (Steady · fresh release) + + Adoption tier is the percentile of total downloads within this collection; + momentum is the percentile of downloads-per-day since release. Momentum is + ranked rather than measured against a fixed rate, because this base model + is itself new and every LoRA looks fast on an absolute scale. Both numbers + are CivitAI's, read on the scrape date at the foot of this file. + +─────────────────────────────────────────────────────────────────────────────── + CHAINING / STACKING +─────────────────────────────────────────────────────────────────────────────── + Stacks with others not stated + + The author gave no chaining guidance. The rules below are the + general Krea 2 ones, not anything specific to this LoRA. + + General Krea 2 rules that apply here (see REFERENCE.txt): + + • The most important rule is not about chaining but about which + checkpoint you chain onto: the official README says in bold "TRAIN on + Raw and RUN on Turbo", and "We highly recommend using RAW for training + LoRAs and applying them on Turbo for inference." LoRAs trained on Raw + are stated to work well on Turbo. + • Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 + LoRAs are DiT-only, so there is no CLIP side to chain and the text- + encoder branch stays untouched however many LoRAs you add. + • Krea 2 LoRAs stack in practice and the community treats multi-LoRA as + routine - style + character, style + slider, realism + character. + Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one + writes "I recently stacked over 20 of my own LoRAs on a single + generation and the output still came out great." + • Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge + on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is + loaded. The CivitAI 12GB training guide also notes rank affects + composability: "rank 8 produced a lighter style influence that was + easier to stack with other LoRAs, rank 16 provided a useful balance + between concept strength and composability, and rank 32 was better + suited to a precise object or concept intended to take priority over + other stacked LoRAs." + • Accelerator LoRAs are position-sensitive in intent, not in node order: + a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a + sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to- + Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is + explicitly recommended by the 4-step LoRA's author. + • Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI + (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card + says "Keep mu = 1.15. The training targets are anchored to that grid; a + different shift moves the [sampler off them]"; the same reasoning + applies to any Turbo-trained adapter in the chain. + • Reference/edit LoRAs are designed to be stacked under ordinary subject + and style LoRAs - the Identity Edit card states "Composes with your + LoRAs: character/body/style LoRAs stack on top and steer the prior" and + recommends a stacked subject LoRA when unusual subjects drift toward + the base prior. + • Krea 2 LoRAs cannot be chained with LoRAs for any other base model. + Many CivitAI pages carry multi-model titles like "(Krea-2 + ZIT)" or + "[Flux | ZIB | Krea2]" - those are separate files for separate + architectures shipped under one model page, not a combined adapter. + • To combine two full Krea 2 checkpoints rather than adapters, ComfyUI + ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per- + section weights for first., tmlp., txtmlp., tproj., + txtfusion.projector, txtfusion.layerwise_blocks.0-1, + txtfusion.refiner_blocks.0-1, blocks.0-27 and last. + • The author never states which weight family this targets. Silent + degradation, never a hard error, between Raw and Turbo. The two share + one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" + - so every LoRA tensor shape matches on both and ComfyUI/diffuser + +─────────────────────────────────────────────────────────────────────────────── + PROS +─────────────────────────────────────────────────────────────────────────────── + + Well established — 4,912 downloads. [source: CivitAI stats] + + Positively received — 290 thumbs-up with no down-votes. [source: + CivitAI stats] + + Explicit trigger word(s) — Digital illustration of, Digital artwork of, + Fantasy, Cartoon — 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 +─────────────────────────────────────────────────────────────────────────────── + [BSS] – Styles for Krea2 - 200📥= New Style ✅ + + - 500📥 = New Style ✅ + + - 1000📥 = New Style ✅ + + - 2000📥 = New Style ✅ + + - 4000📥 = New Style ✅ + + - 6000 📥 = New Style ⏳ + + Vote for the next style in Boosty or Telegram + This pack contains a set of style LoRAs trained specifically on Krea2 . + Each style was designed to expand creative possibilities , giving you more + flexibility when working with Krea2 — whether you want subtle stylistic + shifts or stronger artistic transformations. + + ❤️ Support & Feedback If you enjoy using Equinox , please: 👍 Like the model + 📤 Share your generations 💬 Leave a comment 💡 Suggest ideas for future + versions ☕ Support development via Boosty + https://t.me/blacksnowskill - news + +─────────────────────────────────────────────────────────────────────────────── + 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, 840x1256) + 02_image.jpeg (image, 840x1256) + 03_image.jpeg (image, 840x1256) + 04_image.jpeg (image, NonexNone) + 05_image.jpeg (image, 840x1256) + 06_image.jpeg (image, NonexNone) + +─────────────────────────────────────────────────────────────────────────────── + 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 00:18 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/bss-styles-for-krea2/metadata.json b/bss-styles-for-krea2/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..29137ec7dd65414604ae72356bb190cc142f19f0 --- /dev/null +++ b/bss-styles-for-krea2/metadata.json @@ -0,0 +1,135 @@ +{ + "name": "[BSS] - Styles for Krea2", + "slug": "bss-styles-for-krea2", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 2762978, + "version_id": 3268131, + "version_name": "Digital illustration", + "source": "civitai", + "url": "https://civitai.com/models/2762978", + "repo_id": null, + "file_path": null, + "author": "llikswonskcalb", + "category": { + "key": "style", + "label": "Visual style" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [ + "Digital illustration of", + "Digital artwork of", + "Fantasy", + "Cartoon" + ], + "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": [ + "Well established — 4,912 downloads. [source: CivitAI stats]", + "Positively received — 290 thumbs-up with no down-votes. [source: CivitAI stats]", + "Explicit trigger word(s) — Digital illustration of, Digital artwork of, Fantasy, Cartoon — 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": "DD_krea2_3265864_epoch_4.safetensors", + "size_bytes": 228588904, + "size_mb": 218.0, + "sha256": "ac64c446376ed317dc4c2215c1b8d6e1790b860b78d4361176c29da624ee7080" + }, + "stats": { + "downloads": 4912, + "thumbs_up": 290, + "thumbs_down": 0 + }, + "popularity": { + "downloads": 4912, + "thumbs_up": 290, + "thumbs_down": 0, + "comments": 0, + "approval": "100% positive (290/290)", + "age_days": 9.7, + "downloads_per_day": 35.8, + "rank_by_downloads": 118, + "rank_by_momentum": 65, + "of_total": 150, + "adoption_tier": "Niche", + "momentum": "Steady · fresh release" + }, + "selection_score": 8792.0, + "selection_reasons": [ + "4,912 downloads", + "290 thumbs-up", + "documented triggers + usage notes", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "840x1256", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "840x1256", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "840x1256", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "NonexNone", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "840x1256", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "NonexNone", + "prompt": "" + } + ], + "tags": [ + "style" + ], + "scraped_at": "2026-09-05T00:18:21+00:00", + "path": "krea-2/bss-styles-for-krea2" +} diff --git a/canon-ultrareal/canon_krea2.safetensors b/canon-ultrareal/canon_krea2.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..4d5fc1eac0b63d654551dfff410fd0a25d2e663c --- /dev/null +++ b/canon-ultrareal/canon_krea2.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3295dec59ab1195631fe9b3dd3493ba9c1546056da86179cc3119c4b029420ce +size 114324728 diff --git a/canon-ultrareal/example_images/01_image.jpeg b/canon-ultrareal/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..1894f31e074c05c4b4187059edd6dc4f3af65038 --- /dev/null +++ b/canon-ultrareal/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6877a62e3471741db60055091178d6e4b675268644af291422d8df5b138380f8 +size 7493846 diff --git a/canon-ultrareal/example_images/02_image.jpeg b/canon-ultrareal/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..78fd0fb1214a477b94b395dfe44b0678715ea0b4 --- /dev/null +++ b/canon-ultrareal/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0a7169dd6716d96c0ada6640932a77e22d824027a6bfe6543527dddf97385723 +size 2480485 diff --git a/canon-ultrareal/example_images/03_image.jpeg b/canon-ultrareal/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..9ab1b8d4b73ab8234ea0438d4be79193db0e9e7d --- /dev/null +++ b/canon-ultrareal/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f457937cbc6e3bebaef88054a0af062d2e16bfd36ce2a5b32513f3093bb7e52 +size 2093922 diff --git a/canon-ultrareal/example_images/04_image.jpeg b/canon-ultrareal/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..e39e2b836dc9d03ade2add7da7de829e5dd4404c --- /dev/null +++ b/canon-ultrareal/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d7f0b55a149bb5721e987d51e5b80790ecbe964e0d106fbb8c5fc0d594c84f2e +size 4359538 diff --git a/canon-ultrareal/example_images/05_image.jpeg b/canon-ultrareal/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..5e3a5a9d2709486dd88dc61efa294aa25ae6487b --- /dev/null +++ b/canon-ultrareal/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9496adc36e9609c808d8b02328c4b0d50f224a9cf1ab24a39ccd91e44aba1ad9 +size 1728611 diff --git a/canon-ultrareal/example_images/06_image.jpeg b/canon-ultrareal/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..c566287a67ba5fb363eb72034090a8caa835e880 --- /dev/null +++ b/canon-ultrareal/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ae604232e366704140e1b155225f232c2169b39528a1fd28560d0f304e041c7 +size 2451835 diff --git a/canon-ultrareal/info.txt b/canon-ultrareal/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..f180ec24602743005fcbe99af97f9a38cd0c6518 --- /dev/null +++ b/canon-ultrareal/info.txt @@ -0,0 +1,279 @@ +═══════════════════════════════════════════════════════════════════════════════ + CANON ULTRAREAL +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Photorealism + Author Danrisi + Source https://civitai.com/models/2783143 + Provenance CivitAI · model 2783143 · version 3134717 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File canon_krea2.safetensors + Version v1.0 Krea2 + Size 109.0 MB + Format SafeTensor / unknown precision + SHA256 3295dec59ab1195631fe9b3dd3493ba9c1546056da86179cc3119c4b029420ce + Published 2026-07-16 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + › "c2n0n" + +─────────────────────────────────────────────────────────────────────────────── + 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 5,033 + Rating 100% positive (302/302) + Comments 0 + Published 51 days ago + Download rate 99.6/day since release + + Adoption rank #115 of 150 in this collection (Niche) + Momentum rank #26 of 150 by download rate (Trending) + + Adoption tier is the percentile of total downloads within this collection; + momentum is the percentile of downloads-per-day since release. Momentum is + ranked rather than measured against a fixed rate, because this base model + is itself new and every LoRA looks fast on an absolute scale. Both numbers + are CivitAI's, read on the scrape date at the foot of this file. + +─────────────────────────────────────────────────────────────────────────────── + CHAINING / STACKING +─────────────────────────────────────────────────────────────────────────────── + Stacks with others not stated + + The author gave no chaining guidance. The rules below are the + general Krea 2 ones, not anything specific to this LoRA. + + General Krea 2 rules that apply here (see REFERENCE.txt): + + • The most important rule is not about chaining but about which + checkpoint you chain onto: the official README says in bold "TRAIN on + Raw and RUN on Turbo", and "We highly recommend using RAW for training + LoRAs and applying them on Turbo for inference." LoRAs trained on Raw + are stated to work well on Turbo. + • Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 + LoRAs are DiT-only, so there is no CLIP side to chain and the text- + encoder branch stays untouched however many LoRAs you add. + • Krea 2 LoRAs stack in practice and the community treats multi-LoRA as + routine - style + character, style + slider, realism + character. + Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one + writes "I recently stacked over 20 of my own LoRAs on a single + generation and the output still came out great." + • Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge + on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is + loaded. The CivitAI 12GB training guide also notes rank affects + composability: "rank 8 produced a lighter style influence that was + easier to stack with other LoRAs, rank 16 provided a useful balance + between concept strength and composability, and rank 32 was better + suited to a precise object or concept intended to take priority over + other stacked LoRAs." + • Accelerator LoRAs are position-sensitive in intent, not in node order: + a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a + sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to- + Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is + explicitly recommended by the 4-step LoRA's author. + • Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI + (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card + says "Keep mu = 1.15. The training targets are anchored to that grid; a + different shift moves the [sampler off them]"; the same reasoning + applies to any Turbo-trained adapter in the chain. + • Reference/edit LoRAs are designed to be stacked under ordinary subject + and style LoRAs - the Identity Edit card states "Composes with your + LoRAs: character/body/style LoRAs stack on top and steer the prior" and + recommends a stacked subject LoRA when unusual subjects drift toward + the base prior. + • Krea 2 LoRAs cannot be chained with LoRAs for any other base model. + Many CivitAI pages carry multi-model titles like "(Krea-2 + ZIT)" or + "[Flux | ZIB | Krea2]" - those are separate files for separate + architectures shipped under one model page, not a combined adapter. + • To combine two full Krea 2 checkpoints rather than adapters, ComfyUI + ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per- + section weights for first., tmlp., txtmlp., tproj., + txtfusion.projector, txtfusion.layerwise_blocks.0-1, + txtfusion.refiner_blocks.0-1, blocks.0-27 and last. + • The author never states which weight family this targets. Silent + degradation, never a hard error, between Raw and Turbo. The two share + one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" + - so every LoRA tensor shape matches on both and ComfyUI/diffuser + +─────────────────────────────────────────────────────────────────────────────── + PROS +─────────────────────────────────────────────────────────────────────────────── + + Widely adopted — 5,033 downloads. [source: CivitAI stats] + + Positively received — 302 thumbs-up with no down-votes. [source: + CivitAI stats] + + Explicit trigger word(s) — c2n0n — 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] + +─────────────────────────────────────────────────────────────────────────────── + 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 was trained on my photographs taken with the Canon 1Ds. It is + mainly designed to reproduce its rich colors, distinctive rendering, and + beautiful bokeh that makes old flagmanship full-frame camera. The dataset + includes images shot with 28mm, 50mm, 100mm, and 500mm lenses, so the + resulting look may vary depending on the subject, composition, and focal + length Prompt should be detailed (trigger word is not enough), try to + describe about out-of-focus objects on foreground or background + +─────────────────────────────────────────────────────────────────────────────── + 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, 1536x2728) + 02_image.jpeg (image, 1216x1664) + 03_image.jpeg (image, 1664x1216) + 04_image.jpeg (image, 1088x1928) + 05_image.jpeg (image, 1664x1216) + 06_image.jpeg (image, 1152x1536) + +─────────────────────────────────────────────────────────────────────────────── + 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 00:17 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/canon-ultrareal/metadata.json b/canon-ultrareal/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..e1ee22b8f17b4febcd3a021acea7b1a4b1fbd32b --- /dev/null +++ b/canon-ultrareal/metadata.json @@ -0,0 +1,132 @@ +{ + "name": "Canon UltraReal", + "slug": "canon-ultrareal", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 2783143, + "version_id": 3134717, + "version_name": "v1.0 Krea2", + "source": "civitai", + "url": "https://civitai.com/models/2783143", + "repo_id": null, + "file_path": null, + "author": "Danrisi", + "category": { + "key": "photoreal", + "label": "Photorealism" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [ + "c2n0n" + ], + "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 — 5,033 downloads. [source: CivitAI stats]", + "Positively received — 302 thumbs-up with no down-votes. [source: CivitAI stats]", + "Explicit trigger word(s) — c2n0n — 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]" + ], + "cons": [ + "Author never states which generation mode it was trained for; test against your own workflow before relying on it. [source: description]" + ], + "file": { + "name": "canon_krea2.safetensors", + "size_bytes": 114324728, + "size_mb": 109.03, + "sha256": "3295dec59ab1195631fe9b3dd3493ba9c1546056da86179cc3119c4b029420ce" + }, + "stats": { + "downloads": 5033, + "thumbs_up": 302, + "thumbs_down": 0 + }, + "popularity": { + "downloads": 5033, + "thumbs_up": 302, + "thumbs_down": 0, + "comments": 0, + "approval": "100% positive (302/302)", + "age_days": 50.5, + "downloads_per_day": 99.6, + "rank_by_downloads": 115, + "rank_by_momentum": 26, + "of_total": 150, + "adoption_tier": "Niche", + "momentum": "Trending" + }, + "selection_score": 9057.0, + "selection_reasons": [ + "5,033 downloads", + "302 thumbs-up", + "documented triggers + usage notes", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "1536x2728", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "1216x1664", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "1664x1216", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "1088x1928", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "1664x1216", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "1152x1536", + "prompt": "" + } + ], + "tags": [ + "style" + ], + "scraped_at": "2026-09-05T00:17:51+00:00", + "path": "krea-2/canon-ultrareal" +} diff --git a/cyaniji-niji-style-anima-krea2-nai-zit/Cyaniji Krea2 v1 EP5.safetensors b/cyaniji-niji-style-anima-krea2-nai-zit/Cyaniji Krea2 v1 EP5.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..87092f55d592f9ba2df99174bdf325e28123559b --- /dev/null +++ b/cyaniji-niji-style-anima-krea2-nai-zit/Cyaniji Krea2 v1 EP5.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8bf0f22806cd48f9236be08a334946d962510cb49edbc9fc50b8af932bed102c +size 234703776 diff --git a/cyaniji-niji-style-anima-krea2-nai-zit/example_images/01_image.jpeg b/cyaniji-niji-style-anima-krea2-nai-zit/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..eb11dc276a2e93fe01a0a48e6f496ce0343c2c74 --- /dev/null +++ b/cyaniji-niji-style-anima-krea2-nai-zit/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0839bef7ad8a0c77a5b4998789e7cdbfd581afa1b33b4514f0dcfe55dd2e4541 +size 3277116 diff --git a/cyaniji-niji-style-anima-krea2-nai-zit/example_images/02_image.jpeg b/cyaniji-niji-style-anima-krea2-nai-zit/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..dedbc757211fa654c595a31586ed5499dcdfbb33 --- /dev/null +++ b/cyaniji-niji-style-anima-krea2-nai-zit/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e719d85d0d6d249e946f4d076352b50f7e7b15ccfd84880afbdfe2369d052d52 +size 2239509 diff --git a/cyaniji-niji-style-anima-krea2-nai-zit/example_images/03_image.jpeg b/cyaniji-niji-style-anima-krea2-nai-zit/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..47b7ced52b509968a4926dd95cce423363337614 --- /dev/null +++ b/cyaniji-niji-style-anima-krea2-nai-zit/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e3a9c4f3d9e8047964453f6bd3a4f70b28503c0a89229d80dddc6975c6c1cd95 +size 242767 diff --git a/cyaniji-niji-style-anima-krea2-nai-zit/example_images/04_image.jpeg b/cyaniji-niji-style-anima-krea2-nai-zit/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..08597b1d351ae32fbf596e9b5905c8931d5b6833 --- /dev/null +++ b/cyaniji-niji-style-anima-krea2-nai-zit/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fcf1ee92c7689b08ec5612468535ec425cf671d40c6a5e92f39fab6c204cbdad +size 2034955 diff --git a/cyaniji-niji-style-anima-krea2-nai-zit/example_images/05_image.jpeg b/cyaniji-niji-style-anima-krea2-nai-zit/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..d7b5b471a82a3602c6efc123f7e70b83c50baa6d --- /dev/null +++ b/cyaniji-niji-style-anima-krea2-nai-zit/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b9a064aad41055b08daac922715d3a23593e0bb745b3255836c5f013fb9dd256 +size 3572168 diff --git a/cyaniji-niji-style-anima-krea2-nai-zit/example_images/06_image.jpeg b/cyaniji-niji-style-anima-krea2-nai-zit/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..f5222f4e25fa5508091fa9badea5f8e1bf70a552 --- /dev/null +++ b/cyaniji-niji-style-anima-krea2-nai-zit/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1d9b000916fd745cd860214a1ad285d58c423c6fc291a9783dbefbbfda67fbd8 +size 3899101 diff --git a/cyaniji-niji-style-anima-krea2-nai-zit/info.txt b/cyaniji-niji-style-anima-krea2-nai-zit/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..cdfaed2aabb2f58d9eaf7d1c1c0a18886d0b56e5 --- /dev/null +++ b/cyaniji-niji-style-anima-krea2-nai-zit/info.txt @@ -0,0 +1,252 @@ +═══════════════════════════════════════════════════════════════════════════════ + CYANIJI | NIJI STYLE [ANIMA/KREA2/NAI/ZIT] +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Visual style + Author Imagine__That + Source https://civitai.com/models/2332845 + Provenance CivitAI · model 2332845 · version 3129745 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File Cyaniji Krea2 v1 EP5.safetensors + Version Krea 2 v1.0 + Size 223.8 MB + Format SafeTensor / unknown precision + SHA256 8bf0f22806cd48f9236be08a334946d962510cb49edbc9fc50b8af932bed102c + Published 2026-07-14 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + › "Cyaniji" + +─────────────────────────────────────────────────────────────────────────────── + RECOMMENDED SETTINGS +─────────────────────────────────────────────────────────────────────────────── + Taken verbatim from the author's description: + + LoRA strength 0.3 + +─────────────────────────────────────────────────────────────────────────────── + ADOPTION & POPULARITY +─────────────────────────────────────────────────────────────────────────────── + Downloads 3,998 + Rating 100% positive (388/388) + Comments 0 + Published 52 days ago + Download rate 13.4/day since release + + Adoption rank #132 of 150 in this collection (Niche) + Momentum rank #122 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 +─────────────────────────────────────────────────────────────────────────────── + + Well established — 3,998 downloads. [source: CivitAI stats] + + Positively received — 388 thumbs-up with no down-votes. [source: + CivitAI stats] + + Explicit trigger word(s) — Cyaniji — so the effect can be turned on and + off from the prompt. [source: model metadata] + + Compact at 224 MB — cheap to stack with other LoRAs. [source: file + size] + + Author published concrete recommended settings. [source: description] + +─────────────────────────────────────────────────────────────────────────────── + CONS & CAVEATS +─────────────────────────────────────────────────────────────────────────────── + - Author labels this experimental — results vary shot to shot. [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 +─────────────────────────────────────────────────────────────────────────────── + Trained on my Niji v7 gens. Creates images with a cool tint and vivid + lighting. Does mature men pretty well as well! + Trigger word: Cyaniji + Suggested strength of 0.3-1 depending on how many Loras you're using. + Trained on NAI EPS but works well with most base models. ZIT version is + experimental. Might struggle with faces when it comes to full body/wide + shots. Suggested strength: 0.8-1 + +─────────────────────────────────────────────────────────────────────────────── + 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, 1200x1600) + 02_image.jpeg (image, 1024x1536) + 03_image.jpeg (image, 928x1152) + 04_image.jpeg (image, 1024x1536) + 05_image.jpeg (image, 2730x4096) + 06_image.jpeg (image, 2730x4096) + +─────────────────────────────────────────────────────────────────────────────── + 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 00:18 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/cyaniji-niji-style-anima-krea2-nai-zit/metadata.json b/cyaniji-niji-style-anima-krea2-nai-zit/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..32e87ef001f5c62f3f070461e020b61546bf75ba --- /dev/null +++ b/cyaniji-niji-style-anima-krea2-nai-zit/metadata.json @@ -0,0 +1,140 @@ +{ + "name": "Cyaniji | Niji Style [Anima/Krea2/NAI/ZIT]", + "slug": "cyaniji-niji-style-anima-krea2-nai-zit", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 2332845, + "version_id": 3129745, + "version_name": "Krea 2 v1.0", + "source": "civitai", + "url": "https://civitai.com/models/2332845", + "repo_id": null, + "file_path": null, + "author": "Imagine__That", + "category": { + "key": "style", + "label": "Visual style" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [ + "Cyaniji" + ], + "recommended_settings": { + "LoRA strength": "0.3" + }, + "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": [ + "Well established — 3,998 downloads. [source: CivitAI stats]", + "Positively received — 388 thumbs-up with no down-votes. [source: CivitAI stats]", + "Explicit trigger word(s) — Cyaniji — so the effect can be turned on and off from the prompt. [source: model metadata]", + "Compact at 224 MB — cheap to stack with other LoRAs. [source: file size]", + "Author published concrete recommended settings. [source: description]" + ], + "cons": [ + "Author labels this experimental — results vary shot to shot. [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": "Cyaniji Krea2 v1 EP5.safetensors", + "size_bytes": 234703776, + "size_mb": 223.83, + "sha256": "8bf0f22806cd48f9236be08a334946d962510cb49edbc9fc50b8af932bed102c" + }, + "stats": { + "downloads": 3998, + "thumbs_up": 388, + "thumbs_down": 0 + }, + "popularity": { + "downloads": 3998, + "thumbs_up": 388, + "thumbs_down": 0, + "comments": 0, + "approval": "100% positive (388/388)", + "age_days": 52.3, + "downloads_per_day": 13.38, + "rank_by_downloads": 132, + "rank_by_momentum": 122, + "of_total": 150, + "adoption_tier": "Niche", + "momentum": "Long tail" + }, + "selection_score": 9054.0, + "selection_reasons": [ + "3,998 downloads", + "388 thumbs-up", + "documented triggers + usage notes", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "1200x1600", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "1024x1536", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "928x1152", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "1024x1536", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "2730x4096", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "2730x4096", + "prompt": "" + } + ], + "tags": [ + "style", + "lighting", + "midjourney", + "niji", + "semi realistic" + ], + "scraped_at": "2026-09-05T00:18:00+00:00", + "path": "krea-2/cyaniji-niji-style-anima-krea2-nai-zit" +} diff --git a/dark-fantasy-pulp-pinup/DarkFantasyPulpPinupKrea2Raw.safetensors b/dark-fantasy-pulp-pinup/DarkFantasyPulpPinupKrea2Raw.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..e1e55517fc133dae461ff9ce286f51ca8106a259 --- /dev/null +++ b/dark-fantasy-pulp-pinup/DarkFantasyPulpPinupKrea2Raw.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b2dfc536c5f0df55f6cce3e4653e25ff68e97cd0d2b5d232034f185983cb556 +size 228587752 diff --git a/dark-fantasy-pulp-pinup/example_images/01_image.jpeg b/dark-fantasy-pulp-pinup/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..1a233ac4a83e4e75f47722b4f1b87692720ad329 --- /dev/null +++ b/dark-fantasy-pulp-pinup/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8bbb570fe509e3c9adb55382854483dff7f4a3ab11fc1b5d18e18e6ef882c3c3 +size 1488126 diff --git a/dark-fantasy-pulp-pinup/example_images/02_image.jpeg b/dark-fantasy-pulp-pinup/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..09f0305680493ef36c995b005fd8e7367097c75c --- /dev/null +++ b/dark-fantasy-pulp-pinup/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:61ff472edcc2dc0b2e238975db521b59cdbc227eab21c2a4c47bcbf4c6396cea +size 1716946 diff --git a/dark-fantasy-pulp-pinup/example_images/03_image.jpeg b/dark-fantasy-pulp-pinup/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..7f1ecbb848e6e704a415a3cd631a7e7ca9be7979 --- /dev/null +++ b/dark-fantasy-pulp-pinup/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b4811d48f7649c46ca9ee9a254d2fc14b24b65be7cc0ac02f84af624e6371296 +size 1292052 diff --git a/dark-fantasy-pulp-pinup/example_images/04_image.jpeg b/dark-fantasy-pulp-pinup/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..2b8ca7317a42e83dcdc986d8fa3cb04496017785 --- /dev/null +++ b/dark-fantasy-pulp-pinup/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:26e91c6a0a7bbad9d08de58a3d972827f62eb2b9aa0c4616abf66302e667107f +size 1391212 diff --git a/dark-fantasy-pulp-pinup/example_images/05_image.jpeg b/dark-fantasy-pulp-pinup/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..2a64ad36d2400b621690b42df8a2ef79a9dff70b --- /dev/null +++ b/dark-fantasy-pulp-pinup/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0040d417747371435f126b857a6abe27857c35adc7707d1dfb71754f78102c2d +size 1718313 diff --git a/dark-fantasy-pulp-pinup/example_images/06_image.jpeg b/dark-fantasy-pulp-pinup/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..5e19ab9f79eb5fbb8444332fc8241d87496f0823 --- /dev/null +++ b/dark-fantasy-pulp-pinup/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ac5591814b3f24ab136174b0cf5815a236bb6209ab51e101f0d2bc5c7f77d7a1 +size 1554040 diff --git a/dark-fantasy-pulp-pinup/info.txt b/dark-fantasy-pulp-pinup/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..e72d136e17a11647a2633ede125bccda177a390d --- /dev/null +++ b/dark-fantasy-pulp-pinup/info.txt @@ -0,0 +1,277 @@ +═══════════════════════════════════════════════════════════════════════════════ + DARK FANTASY PULP PINUP +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + 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/1550211 + Provenance CivitAI · model 1550211 · version 3166814 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File DarkFantasyPulpPinupKrea2Raw.safetensors + Version Krea2 + Size 218.0 MB + Format SafeTensor / unknown precision + SHA256 0b2dfc536c5f0df55f6cce3e4653e25ff68e97cd0d2b5d232034f185983cb556 + Published 2026-08-18 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + › "Dark Fantasy Pulp Pinup," + +─────────────────────────────────────────────────────────────────────────────── + 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 4,556 + Rating 100% positive (370/370) + Comments 0 + Published 18 days ago + Download rate 31.3/day since release + + Adoption rank #125 of 150 in this collection (Niche) + Momentum rank #75 of 150 by download rate (Steady · fresh release) + + Adoption tier is the percentile of total downloads within this collection; + momentum is the percentile of downloads-per-day since release. Momentum is + ranked rather than measured against a fixed rate, because this base model + is itself new and every LoRA looks fast on an absolute scale. Both numbers + are CivitAI's, read on the scrape date at the foot of this file. + +─────────────────────────────────────────────────────────────────────────────── + CHAINING / STACKING +─────────────────────────────────────────────────────────────────────────────── + Stacks with others not stated + + The author gave no chaining guidance. The rules below are the + general Krea 2 ones, not anything specific to this LoRA. + + General Krea 2 rules that apply here (see REFERENCE.txt): + + • The most important rule is not about chaining but about which + checkpoint you chain onto: the official README says in bold "TRAIN on + Raw and RUN on Turbo", and "We highly recommend using RAW for training + LoRAs and applying them on Turbo for inference." LoRAs trained on Raw + are stated to work well on Turbo. + • Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 + LoRAs are DiT-only, so there is no CLIP side to chain and the text- + encoder branch stays untouched however many LoRAs you add. + • Krea 2 LoRAs stack in practice and the community treats multi-LoRA as + routine - style + character, style + slider, realism + character. + Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one + writes "I recently stacked over 20 of my own LoRAs on a single + generation and the output still came out great." + • Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge + on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is + loaded. The CivitAI 12GB training guide also notes rank affects + composability: "rank 8 produced a lighter style influence that was + easier to stack with other LoRAs, rank 16 provided a useful balance + between concept strength and composability, and rank 32 was better + suited to a precise object or concept intended to take priority over + other stacked LoRAs." + • Accelerator LoRAs are position-sensitive in intent, not in node order: + a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a + sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to- + Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is + explicitly recommended by the 4-step LoRA's author. + • Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI + (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card + says "Keep mu = 1.15. The training targets are anchored to that grid; a + different shift moves the [sampler off them]"; the same reasoning + applies to any Turbo-trained adapter in the chain. + • Reference/edit LoRAs are designed to be stacked under ordinary subject + and style LoRAs - the Identity Edit card states "Composes with your + LoRAs: character/body/style LoRAs stack on top and steer the prior" and + recommends a stacked subject LoRA when unusual subjects drift toward + the base prior. + • Krea 2 LoRAs cannot be chained with LoRAs for any other base model. + Many CivitAI pages carry multi-model titles like "(Krea-2 + ZIT)" or + "[Flux | ZIB | Krea2]" - those are separate files for separate + architectures shipped under one model page, not a combined adapter. + • To combine two full Krea 2 checkpoints rather than adapters, ComfyUI + ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per- + section weights for first., tmlp., txtmlp., tproj., + txtfusion.projector, txtfusion.layerwise_blocks.0-1, + txtfusion.refiner_blocks.0-1, blocks.0-27 and last. + • The author never states which weight family this targets. Silent + degradation, never a hard error, between Raw and Turbo. The two share + one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" + - so every LoRA tensor shape matches on both and ComfyUI/diffuser + +─────────────────────────────────────────────────────────────────────────────── + PROS +─────────────────────────────────────────────────────────────────────────────── + + Well established — 4,556 downloads. [source: CivitAI stats] + + Positively received — 370 thumbs-up with no down-votes. [source: + CivitAI stats] + + Explicit trigger word(s) — Dark Fantasy Pulp Pinup, — 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 +─────────────────────────────────────────────────────────────────────────────── + Classic pin-up style infused with dark fantasy flair, this LoRA draws from + the energy of the 80s pulp art. + Expect striking figures, dramatic poses, and rich, moody atmospheres. + Inspired by the iconic works of Olivia De Berardinis , Luis Royo , and + Frank Frazetta . + +─────────────────────────────────────────────────────────────────────────────── + 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, 1280x720) + 02_image.jpeg (image, 720x1280) + 03_image.jpeg (image, 768x1024) + 04_image.jpeg (image, 720x1280) + 05_image.jpeg (image, 720x1280) + 06_image.jpeg (image, 1280x720) + +─────────────────────────────────────────────────────────────────────────────── + 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 00:18 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/dark-fantasy-pulp-pinup/metadata.json b/dark-fantasy-pulp-pinup/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..c17d1dec98fa2f191af1f083e9d7b2d369f8f25a --- /dev/null +++ b/dark-fantasy-pulp-pinup/metadata.json @@ -0,0 +1,136 @@ +{ + "name": "Dark Fantasy Pulp Pinup", + "slug": "dark-fantasy-pulp-pinup", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 1550211, + "version_id": 3166814, + "version_name": "Krea2", + "source": "civitai", + "url": "https://civitai.com/models/1550211", + "repo_id": null, + "file_path": null, + "author": "ArsMachina", + "category": { + "key": "style", + "label": "Visual style" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [ + "Dark Fantasy Pulp Pinup," + ], + "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": [ + "Well established — 4,556 downloads. [source: CivitAI stats]", + "Positively received — 370 thumbs-up with no down-votes. [source: CivitAI stats]", + "Explicit trigger word(s) — Dark Fantasy Pulp Pinup, — 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": "DarkFantasyPulpPinupKrea2Raw.safetensors", + "size_bytes": 228587752, + "size_mb": 218.0, + "sha256": "0b2dfc536c5f0df55f6cce3e4653e25ff68e97cd0d2b5d232034f185983cb556" + }, + "stats": { + "downloads": 4556, + "thumbs_up": 370, + "thumbs_down": 0 + }, + "popularity": { + "downloads": 4556, + "thumbs_up": 370, + "thumbs_down": 0, + "comments": 0, + "approval": "100% positive (370/370)", + "age_days": 17.5, + "downloads_per_day": 31.28, + "rank_by_downloads": 125, + "rank_by_momentum": 75, + "of_total": 150, + "adoption_tier": "Niche", + "momentum": "Steady · fresh release" + }, + "selection_score": 8996.0, + "selection_reasons": [ + "4,556 downloads", + "370 thumbs-up", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "1280x720", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "720x1280", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "768x1024", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "720x1280", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "720x1280", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "1280x720", + "prompt": "" + } + ], + "tags": [ + "style", + "olivia de berardinis", + "dark fantasy", + "frank frazetta", + "pin-up", + "luis royo" + ], + "scraped_at": "2026-09-05T00:18:02+00:00", + "path": "krea-2/dark-fantasy-pulp-pinup" +} diff --git a/delken-furry-realism-enhancer-krea-2-anima-illustrious/Delken_Krea2_V10.safetensors b/delken-furry-realism-enhancer-krea-2-anima-illustrious/Delken_Krea2_V10.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..e06f9357465607ccb80073ee2cc40c9015a88f85 --- /dev/null +++ b/delken-furry-realism-enhancer-krea-2-anima-illustrious/Delken_Krea2_V10.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:813e5182e2543f2d38b62bdf936f180ab8c3a35614249459e785cdb9c4ec5aae +size 114329256 diff --git a/delken-furry-realism-enhancer-krea-2-anima-illustrious/example_images/01_image.jpeg b/delken-furry-realism-enhancer-krea-2-anima-illustrious/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..47e677639b6c084d391957af1ad5604b5d9c6ed8 --- /dev/null +++ b/delken-furry-realism-enhancer-krea-2-anima-illustrious/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4739ee2e48b8447804804dcc0c85150f1cbbe8a5b2bbb0c285469ad3ea59edd8 +size 2458157 diff --git a/delken-furry-realism-enhancer-krea-2-anima-illustrious/example_images/02_image.jpeg b/delken-furry-realism-enhancer-krea-2-anima-illustrious/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..860556cf0acdc3f83a5ff0dc18c48b163dbb79bf --- /dev/null +++ b/delken-furry-realism-enhancer-krea-2-anima-illustrious/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cecf488e5f38072be6c35d9ec62a6562c87f976fb422c025b5b51efff60e440b +size 1677845 diff --git a/delken-furry-realism-enhancer-krea-2-anima-illustrious/example_images/03_image.jpeg b/delken-furry-realism-enhancer-krea-2-anima-illustrious/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..3fb43981d890294e9b240d6561b289397f5e75af --- /dev/null +++ b/delken-furry-realism-enhancer-krea-2-anima-illustrious/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:caeacf5e36faf2647591110f4decb90f92e8866dbbdcd18b440190ea1c93a397 +size 2145551 diff --git a/delken-furry-realism-enhancer-krea-2-anima-illustrious/example_images/04_image.jpeg b/delken-furry-realism-enhancer-krea-2-anima-illustrious/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..eeb41f329718a951e478dfe307dddfa8fb861676 --- /dev/null +++ b/delken-furry-realism-enhancer-krea-2-anima-illustrious/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1b1440f9cf0d52503b418d2384c59e021e2f629aa4ba5a1198971c7d8fac4835 +size 2359939 diff --git a/delken-furry-realism-enhancer-krea-2-anima-illustrious/example_images/05_image.jpeg b/delken-furry-realism-enhancer-krea-2-anima-illustrious/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..f7b0303850458413d4c3832488cc46787524e2b2 --- /dev/null +++ b/delken-furry-realism-enhancer-krea-2-anima-illustrious/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1b70d7e74bb2534c14fb946b4204aaa474edae84bd88bc4832b6f5e0a95cfa72 +size 2489941 diff --git a/delken-furry-realism-enhancer-krea-2-anima-illustrious/example_images/06_image.jpeg b/delken-furry-realism-enhancer-krea-2-anima-illustrious/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..ef21c32205f840166402f60da32854cf58456bc0 --- /dev/null +++ b/delken-furry-realism-enhancer-krea-2-anima-illustrious/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d8a5c3d65918fe60fdbbcc6beb371039b5c3d51d2d25a9669b406e4d269f070 +size 183814 diff --git a/delken-furry-realism-enhancer-krea-2-anima-illustrious/info.txt b/delken-furry-realism-enhancer-krea-2-anima-illustrious/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..d794fad991021115a494eb6f9b381472a2725b81 --- /dev/null +++ b/delken-furry-realism-enhancer-krea-2-anima-illustrious/info.txt @@ -0,0 +1,284 @@ +═══════════════════════════════════════════════════════════════════════════════ + DELKEN FURRY REALISM ENHANCER [KREA 2 / ANIMA / ILLUSTRIOUS] +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Detail / anatomy fixer + Author Boerboer + Source https://civitai.com/models/2458864 + Provenance CivitAI · model 2458864 · version 3289747 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File Delken_Krea2_V10.safetensors + Version v1.0 Krea 2 + Size 109.0 MB + Format SafeTensor / unknown precision + SHA256 813e5182e2543f2d38b62bdf936f180ab8c3a35614249459e785cdb9c4ec5aae + Published 2026-09-02 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + › "@delken" + +─────────────────────────────────────────────────────────────────────────────── + RECOMMENDED SETTINGS +─────────────────────────────────────────────────────────────────────────────── + Taken verbatim from the author's description: + + LoRA strength 0.6 + +─────────────────────────────────────────────────────────────────────────────── + ADOPTION & POPULARITY +─────────────────────────────────────────────────────────────────────────────── + Downloads 3,112 + Rating 100% positive (420/420) + Comments 0 + Published 3 days ago + Download rate 130.2/day since release + + Adoption rank #146 of 150 in this collection (Niche) + Momentum rank #17 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 +─────────────────────────────────────────────────────────────────────────────── + + Well established — 3,112 downloads. [source: CivitAI stats] + + Positively received — 420 thumbs-up with no down-votes. [source: + CivitAI stats] + + Explicit trigger word(s) — @delken — 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 says the LoRA may be overtrained — start at low strength. + [source: description] + - Author labels this experimental — results vary shot to shot. [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 +─────────────────────────────────────────────────────────────────────────────── + The aim of this LoRA is not to be a perfect standalone realism LoRA. For + that a dedicated finetune would usually be better. Instead, it's a + potentially useful tool to enhance the realism and fur detailing of your + favourite checkpoint. Krea 2: I was 50/50 on whether I should make this one + and post it, but I've been happy with some of the results. Does realism the + best out of all the versions which is somewhat expected, but it's not + perfect. Anima: First attempt at an Anima version. Anima obviously isn't + meant to be used for realistic images so don't expect pure realism just + yet. Illustrious: I wish I can say V2 is a flat upgrade over V1, but they + both have there perks. Here's a summary. Feel free to test and choose what + works for you. For both I still highly recommend using a face detailer. + v1.0: Intentionally overtrained on a small dataset (mostly wolves and + tigers) to maximize realism, but at the cost of some usability. Basically + high risk, high reward. Use at 0.6 weight. Higher weights can be done but + at your own risk. v2.0: Doubled the dataset size to include more species, + also trained on more reasonable settings. This is the safer option but + isn't as realistic as v1.0 as a result. Use-able up to 1.0 strength and + beyond. The reason the lora has "Delken" in the name is because that's the + AI director who's images predominantly make up the dataset. I would love to + have a more varied dataset but I haven't found anyone who consistently + matches the quality Delken in pure realism IMO. As of v2.0, I've started + adding my own Chatgpt generated images, but I'm still more than open to + suggestions. Delken if you're reading this, thanks for the awesome pics and + sorry that I stole them. Feel free to DM me if you want the name taken off + the lora or for anything else. + +─────────────────────────────────────────────────────────────────────────────── + VERSION NOTES +─────────────────────────────────────────────────────────────────────────────── + Beefed up the dataset a little compared to the Anima V1.0 version. When + Krea 2 first came out, it seemed a realistic Furry LoRA wasn't needed, but + since I realised nude realistic furries is pretty difficult without LoRAs. + Hopefully this LoRA helps a little with that. I'm not perfectly happy with + it, the fur can look matted at times and the species knowledge is quite + narrow. But I've managed to gen a few images I'm happy with so I though it + was still worth posting. It especially looks good when paired with certain + artist style LoRAs like Anhes. + +─────────────────────────────────────────────────────────────────────────────── + 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, 1344x1728) + 02_image.jpeg (image, 1152x1536) + 03_image.jpeg (image, 1088x1456) + 04_image.jpeg (image, 1088x1456) + 05_image.jpeg (image, 1088x1456) + 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 00:18 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/delken-furry-realism-enhancer-krea-2-anima-illustrious/metadata.json b/delken-furry-realism-enhancer-krea-2-anima-illustrious/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..264153e0226b30329b459f5bb53915c5e2ee4511 --- /dev/null +++ b/delken-furry-realism-enhancer-krea-2-anima-illustrious/metadata.json @@ -0,0 +1,146 @@ +{ + "name": "Delken Furry Realism Enhancer [Krea 2 / Anima / Illustrious]", + "slug": "delken-furry-realism-enhancer-krea-2-anima-illustrious", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 2458864, + "version_id": 3289747, + "version_name": "v1.0 Krea 2", + "source": "civitai", + "url": "https://civitai.com/models/2458864", + "repo_id": null, + "file_path": null, + "author": "Boerboer", + "category": { + "key": "detail", + "label": "Detail / anatomy fixer" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [ + "@delken" + ], + "recommended_settings": { + "LoRA strength": "0.6" + }, + "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": [ + "Well established — 3,112 downloads. [source: CivitAI stats]", + "Positively received — 420 thumbs-up with no down-votes. [source: CivitAI stats]", + "Explicit trigger word(s) — @delken — 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 says the LoRA may be overtrained — start at low strength. [source: description]", + "Author labels this experimental — results vary shot to shot. [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": "Delken_Krea2_V10.safetensors", + "size_bytes": 114329256, + "size_mb": 109.03, + "sha256": "813e5182e2543f2d38b62bdf936f180ab8c3a35614249459e785cdb9c4ec5aae" + }, + "stats": { + "downloads": 3112, + "thumbs_up": 420, + "thumbs_down": 0 + }, + "popularity": { + "downloads": 3112, + "thumbs_up": 420, + "thumbs_down": 0, + "comments": 0, + "approval": "100% positive (420/420)", + "age_days": 2.5, + "downloads_per_day": 130.16, + "rank_by_downloads": 146, + "rank_by_momentum": 17, + "of_total": 150, + "adoption_tier": "Niche", + "momentum": "Trending · fresh release" + }, + "selection_score": 8552.0, + "selection_reasons": [ + "3,112 downloads", + "420 thumbs-up", + "documented triggers + usage notes", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "1344x1728", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "1152x1536", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "1088x1456", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "1088x1456", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "1088x1456", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "1184x896", + "prompt": "" + } + ], + "tags": [ + "anthro", + "realistic", + "male", + "style", + "furry", + "realism", + "woman", + "realistic furry", + "detailed fur", + "realistic furries" + ], + "scraped_at": "2026-09-05T00:18:23+00:00", + "path": "krea-2/delken-furry-realism-enhancer-krea-2-anima-illustrious" +} diff --git a/furry-realistic-style/FURRY_REALISTIC_STYLE__Krea_2__krea2_3284944_epoch_10.safetensors b/furry-realistic-style/FURRY_REALISTIC_STYLE__Krea_2__krea2_3284944_epoch_10.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..8ad0aa696d01597009e2fe8fe4333797fc7a00c8 --- /dev/null +++ b/furry-realistic-style/FURRY_REALISTIC_STYLE__Krea_2__krea2_3284944_epoch_10.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:36c0f8d4e80f737fa1530c999f1b4b36d9b462457b1319513ffa0d2c51d8cf7e +size 228588920 diff --git a/furry-realistic-style/example_images/01_image.jpeg b/furry-realistic-style/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..43411b736250d8844702226308ed2e4567ecf009 --- /dev/null +++ b/furry-realistic-style/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b1146f30543809cc983c2820fa1a95a55bc42996f3c93c2bb43b8567e4398d79 +size 262388 diff --git a/furry-realistic-style/example_images/02_image.jpeg b/furry-realistic-style/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..5ce13796491a2a471bdaa0e6b3ad554f8ffe6af2 --- /dev/null +++ b/furry-realistic-style/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6b4c3ba17282c49c6cdb85aa1c91d4312ae8670d07b6d0f1fe64e76c25f5409d +size 257923 diff --git a/furry-realistic-style/example_images/03_image.jpeg b/furry-realistic-style/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..9a6f3a50cab9a827c7f02371d424c3827efd89a2 --- /dev/null +++ b/furry-realistic-style/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0e623a9ad81958c94240b77fc8551f843199ec56661979215188a11a34f8432e +size 194795 diff --git a/furry-realistic-style/example_images/04_image.jpeg b/furry-realistic-style/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..554e45b8baa8e730588279bafe74abf29266f1c0 --- /dev/null +++ b/furry-realistic-style/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5293381d5fc1fde73f90297ae27802fbb34b8c2f2bebfd156f34770e6d05740c +size 283268 diff --git a/furry-realistic-style/example_images/05_image.jpeg b/furry-realistic-style/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..dfc5ada6d947abd4a2073b4203382bd35907f42e --- /dev/null +++ b/furry-realistic-style/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c2f6a2c42725e8246a990e1e4a443e6b55a4119a623041914724e909be1ee7bb +size 230355 diff --git a/furry-realistic-style/example_images/06_image.jpeg b/furry-realistic-style/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..aa5bb50175deb76e76f0fbb00ddd8d8f250834af --- /dev/null +++ b/furry-realistic-style/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:640134d1cc0e79cd9bafc5ed78a0dcff1ffe6f4df09158ab9b8e9c3a0dd8722d +size 325593 diff --git a/furry-realistic-style/info.txt b/furry-realistic-style/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..344c31bd94b1eb052a0e7774503c2be9ee511c1d --- /dev/null +++ b/furry-realistic-style/info.txt @@ -0,0 +1,290 @@ +═══════════════════════════════════════════════════════════════════════════════ + FURRY REALISTIC STYLE +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Photorealism + Author Kayot + Source https://civitai.com/models/936672 + Provenance CivitAI · model 936672 · version 3287711 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File FURRY_REALISTIC_STYLE__Krea_2__krea2_3284944_epoch_10.safetensors + Version FRS (Krea 2) + Size 218.0 MB + Format SafeTensor / unknown precision + SHA256 36c0f8d4e80f737fa1530c999f1b4b36d9b462457b1319513ffa0d2c51d8cf7e + Published 2026-09-01 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + › "R1lisit1ka," + +─────────────────────────────────────────────────────────────────────────────── + 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 3,081 + Rating 100% positive (451/452) + Comments 0 + Published 3 days ago + Download rate 29.3/day since release + + Adoption rank #148 of 150 in this collection (Niche) + Momentum rank #79 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 +─────────────────────────────────────────────────────────────────────────────── + + Well established — 3,081 downloads. [source: CivitAI stats] + + Positively received — 451 thumbs-up, 1 down. [source: CivitAI stats] + + Explicit trigger word(s) — R1lisit1ka, — 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 +─────────────────────────────────────────────────────────────────────────────── + Hi everyone :D This is Furry Realistic Style Lora (Sorry for the cringey + name – a long time ago I thought it would sound cool XD). Basically, the + essence of the style is creating characters in a more or less realistic + atmosphere (I thought plain realism was too boring, so it needed a + distinctive design). Sooo) You can see what it looks like in the sample + images) So... yeah... that's the short version) Have a nice day and + successful generations)))!!! + + Main trigger words: R1lisit1ka, Even though the style name includes + "Furry", it's actually optional – the Lora works with both humans and + animals. There are no special recommendations – just use " R1lisit1ka, " as + the trigger word for the Krea 2 version (earlier versions have different + triggers, but they're listed there). Everything should work intuitively, + and you shouldn't run into any issues. Please feel free to experiment and + share your results))) + + Negative prompts: Watermark, Fused tail, bad eyes, lips, bad anatomy, bad + had, fused fingers, fused face, bad face, bad quality, bad body, fused + legs, bad details, particles, bad clothes, bad hair, fused body, + +─────────────────────────────────────────────────────────────────────────────── + EXAMPLE MEDIA +─────────────────────────────────────────────────────────────────────────────── + 6 preview(s) in ./example_images/ — community results for + this LoRA. Videos keep a matching _poster.jpg still frame. + + No prompts are recorded below: CivitAI's public API flags these posts as + having generation metadata but does not return it, so the prompts cannot be + scraped. Open the model page to read them. + + 01_image.jpeg (image, 832x1248) + 02_image.jpeg (image, 832x1248) + 03_image.jpeg (image, 832x1248) + 04_image.jpeg (image, 832x1248) + 05_image.jpeg (image, 832x1248) + 06_image.jpeg (image, 832x1248) + +─────────────────────────────────────────────────────────────────────────────── + 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 00:18 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/furry-realistic-style/metadata.json b/furry-realistic-style/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..3c697ec0e656286d496b15a0de58337652224abe --- /dev/null +++ b/furry-realistic-style/metadata.json @@ -0,0 +1,137 @@ +{ + "name": "FURRY REALISTIC STYLE", + "slug": "furry-realistic-style", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 936672, + "version_id": 3287711, + "version_name": "FRS (Krea 2)", + "source": "civitai", + "url": "https://civitai.com/models/936672", + "repo_id": null, + "file_path": null, + "author": "Kayot", + "category": { + "key": "photoreal", + "label": "Photorealism" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [ + "R1lisit1ka," + ], + "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": [ + "Well established — 3,081 downloads. [source: CivitAI stats]", + "Positively received — 451 thumbs-up, 1 down. [source: CivitAI stats]", + "Explicit trigger word(s) — R1lisit1ka, — 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": "FURRY_REALISTIC_STYLE__Krea_2__krea2_3284944_epoch_10.safetensors", + "size_bytes": 228588920, + "size_mb": 218.0, + "sha256": "36c0f8d4e80f737fa1530c999f1b4b36d9b462457b1319513ffa0d2c51d8cf7e" + }, + "stats": { + "downloads": 3081, + "thumbs_up": 451, + "thumbs_down": 1 + }, + "popularity": { + "downloads": 3081, + "thumbs_up": 451, + "thumbs_down": 1, + "comments": 0, + "approval": "100% positive (451/452)", + "age_days": 3.2, + "downloads_per_day": 29.3, + "rank_by_downloads": 148, + "rank_by_momentum": 79, + "of_total": 150, + "adoption_tier": "Niche", + "momentum": "Long tail · fresh release" + }, + "selection_score": 8893.0, + "selection_reasons": [ + "3,081 downloads", + "451 thumbs-up", + "documented triggers + usage notes", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "832x1248", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "832x1248", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "832x1248", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "832x1248", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "832x1248", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "832x1248", + "prompt": "" + } + ], + "tags": [ + "realistic", + "woman", + "photorealistic", + "girls", + "furry", + "style" + ], + "scraped_at": "2026-09-05T00:18:06+00:00", + "path": "krea-2/furry-realistic-style" +} diff --git a/hair-length-slider-krea2-zit/example_images/01_image.jpeg b/hair-length-slider-krea2-zit/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..f160e00ffdb8ef5e4fe61258b5c1163c82e0d58b --- /dev/null +++ b/hair-length-slider-krea2-zit/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ce5c5c736f5b4fa2dd0e4de4305a04c0c432e66cb5ab4d71a892a00846a42cd6 +size 288590 diff --git a/hair-length-slider-krea2-zit/example_images/02_image.jpeg b/hair-length-slider-krea2-zit/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..633e035c40ee941f1d1c7f3311b67b310b49ca85 --- /dev/null +++ b/hair-length-slider-krea2-zit/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2e6dfee6d0d937324af8a236fabd56da8cca7e01c6e358f69ee060205c849231 +size 172490 diff --git a/hair-length-slider-krea2-zit/example_images/03_image.jpeg b/hair-length-slider-krea2-zit/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..82334e9999a07b7fd1dca91ef9e8813e21ee04b4 --- /dev/null +++ b/hair-length-slider-krea2-zit/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0bd669afb9174cf2a5c1c78ab89d37aecc32d15499168031b7650f4032464704 +size 4598330 diff --git a/hair-length-slider-krea2-zit/example_images/04_image.jpeg b/hair-length-slider-krea2-zit/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..8940d0d4b6d7bed029082a440b24163c8384a6d2 --- /dev/null +++ b/hair-length-slider-krea2-zit/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8324e22985e87a32aa38322e328684305761efeaab6df3e95410796b7a149422 +size 11630343 diff --git a/hair-length-slider-krea2-zit/example_images/05_image.jpeg b/hair-length-slider-krea2-zit/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..beb8b639f1276115377807cc32f7a0f4582bc75e --- /dev/null +++ b/hair-length-slider-krea2-zit/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:644f827ce061a8dae360f47bc67f0238ee9b78b92f27cf944e3587b439aad39b +size 3680398 diff --git a/hair-length-slider-krea2-zit/example_images/06_image.jpeg b/hair-length-slider-krea2-zit/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..a20acd6ebe2a34dbda025deb8d035f24e7d2fc40 --- /dev/null +++ b/hair-length-slider-krea2-zit/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d5140eaf5973255e38b90f97e02ed5fbbed20c7738baa75344b314454a07a7db +size 5635610 diff --git a/hair-length-slider-krea2-zit/hair_length_krea2_loraholic.safetensors b/hair-length-slider-krea2-zit/hair_length_krea2_loraholic.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..fd0b932b8585e078aba9421cd79c51dfbabb9f3e --- /dev/null +++ b/hair-length-slider-krea2-zit/hair_length_krea2_loraholic.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da9b8202fbb09bfe1a4adbb31e505abfa45aafb1410e9d3caf0491f41f296870 +size 7203000 diff --git a/hair-length-slider-krea2-zit/info.txt b/hair-length-slider-krea2-zit/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..09237116c82de76c463ceb610058aacc091725e3 --- /dev/null +++ b/hair-length-slider-krea2-zit/info.txt @@ -0,0 +1,286 @@ +═══════════════════════════════════════════════════════════════════════════════ + HAIR LENGTH SLIDER - (KREA2 + ZIT) +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + 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/2542224 + Provenance CivitAI · model 2542224 · version 3104273 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File hair_length_krea2_loraholic.safetensors + Version Krea2_v1 + Size 6.9 MB + Format SafeTensor / unknown precision + SHA256 da9b8202fbb09bfe1a4adbb31e505abfa45aafb1410e9d3caf0491f41f296870 + Published 2026-07-06 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + No trigger word. The LoRA applies as soon as it is loaded — + strength is the only way to dial the effect up or down. + +─────────────────────────────────────────────────────────────────────────────── + RECOMMENDED SETTINGS +─────────────────────────────────────────────────────────────────────────────── + 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 5,384 + Rating 100% positive (258/258) + Comments 7 + Published 60 days ago + Download rate 56.6/day since release + + Adoption rank #108 of 150 in this collection (Modest) + Momentum rank #47 of 150 by download rate (Steady) + + 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 — 5,384 downloads. [source: CivitAI stats] + + Positively received — 258 thumbs-up with no down-votes. [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 +─────────────────────────────────────────────────────────────────────────────── + Warning about the Krea2 version. I just found out that it tends to make a + lot of things longer, not just the hair. I will look into this in a v2. But + until i make a v2 i have adjusted the recommended stable range to a range + that should be ok. Outside of "Stable"-range use at your own risk XD + Krea2: Stable from -2.5 to 2.5 (both Krea2 and ZIT) + ZIT: Stable from -4 to 4 But works outside this range as well. Works very + well in combination with my Hair Volume Slider + 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, 832x1248) + 02_image.jpeg (image, 832x1248) + 03_image.jpeg (image, 1792x2432) + 04_image.jpeg (image, 2316x4096) + 05_image.jpeg (image, 1472x2048) + 06_image.jpeg (image, 2048x1792) + +─────────────────────────────────────────────────────────────────────────────── + 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 00:18 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/hair-length-slider-krea2-zit/metadata.json b/hair-length-slider-krea2-zit/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..62ecf67557c3b016b8c0ec4e98313472dd7a9a11 --- /dev/null +++ b/hair-length-slider-krea2-zit/metadata.json @@ -0,0 +1,131 @@ +{ + "name": "Hair Length Slider - (Krea2 + ZIT)", + "slug": "hair-length-slider-krea2-zit", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 2542224, + "version_id": 3104273, + "version_name": "Krea2_v1", + "source": "civitai", + "url": "https://civitai.com/models/2542224", + "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 — 5,384 downloads. [source: CivitAI stats]", + "Positively received — 258 thumbs-up with no down-votes. [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": "hair_length_krea2_loraholic.safetensors", + "size_bytes": 7203000, + "size_mb": 6.87, + "sha256": "da9b8202fbb09bfe1a4adbb31e505abfa45aafb1410e9d3caf0491f41f296870" + }, + "stats": { + "downloads": 5384, + "thumbs_up": 258, + "thumbs_down": 0 + }, + "popularity": { + "downloads": 5384, + "thumbs_up": 258, + "thumbs_down": 0, + "comments": 7, + "approval": "100% positive (258/258)", + "age_days": 60.4, + "downloads_per_day": 56.56, + "rank_by_downloads": 108, + "rank_by_momentum": 47, + "of_total": 150, + "adoption_tier": "Modest", + "momentum": "Steady" + }, + "selection_score": 8480.0, + "selection_reasons": [ + "5,384 downloads", + "258 thumbs-up", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "832x1248", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "832x1248", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "1792x2432", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "2316x4096", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "1472x2048", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "2048x1792", + "prompt": "" + } + ], + "tags": [ + "slider", + "length", + "tool" + ], + "scraped_at": "2026-09-05T00:18:25+00:00", + "path": "krea-2/hair-length-slider-krea2-zit" +} diff --git a/krea-2-weight-slider/WeightSlider-KREA2_v2.safetensors b/krea-2-weight-slider/WeightSlider-KREA2_v2.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..301f148d5d6d7c623a500078b47d6dd897edb53c --- /dev/null +++ b/krea-2-weight-slider/WeightSlider-KREA2_v2.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8145f7315b5eb789bdca76665a5c1ace1f4f9883a1b41c7a02fc574be7234296 +size 7202400 diff --git a/krea-2-weight-slider/example_images/01_image.jpeg b/krea-2-weight-slider/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..d2ec20df5506d1ad373a189406f1cd940d40e764 --- /dev/null +++ b/krea-2-weight-slider/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5490d5d590c29bd32ae93f204dc618495d70947c13e7ce2a2580f333eaaa7a45 +size 474890 diff --git a/krea-2-weight-slider/example_images/02_image.jpeg b/krea-2-weight-slider/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..04fed74aba528a0e5a789dedd8ec22a6ca705c33 --- /dev/null +++ b/krea-2-weight-slider/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cc90fb4fbded7d4d0d441ffa01c192193732604518368742d751c338887bc023 +size 1673506 diff --git a/krea-2-weight-slider/example_images/03_image.jpeg b/krea-2-weight-slider/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..184cb61d0ba5af3c7ae912c21ab8a18993e8bc6d --- /dev/null +++ b/krea-2-weight-slider/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ff6392f7ef4167799ab2ebc5937a76efeb29e8782851a4e5285a88f3c137616 +size 1042076 diff --git a/krea-2-weight-slider/example_images/04_image.jpeg b/krea-2-weight-slider/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..d60dd6aada2ed19276157c84997a1b763133f3eb --- /dev/null +++ b/krea-2-weight-slider/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e49830b6814a8443f39ea8de7ea32e7e89de1b024aba814bd81ce7097ccf601 +size 585513 diff --git a/krea-2-weight-slider/example_images/05_image.jpeg b/krea-2-weight-slider/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..79e9cb4957f11ae0d353fee1404c5d077014dd1c --- /dev/null +++ b/krea-2-weight-slider/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a041b8b8727234441de7a06a187592901e7dd230831a830fb2b1a1f2a3cd3204 +size 720460 diff --git a/krea-2-weight-slider/example_images/06_image.jpeg b/krea-2-weight-slider/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..2a6d7a04e919367b34634962126065ece2fccc82 --- /dev/null +++ b/krea-2-weight-slider/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e0809dacb8f8d9fcdc5d80c128597b43e48f0e05a9a239327f44994248a52772 +size 738213 diff --git a/krea-2-weight-slider/info.txt b/krea-2-weight-slider/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..b9db89099a481c90288ce2cd959d14bc21ad7c1d --- /dev/null +++ b/krea-2-weight-slider/info.txt @@ -0,0 +1,280 @@ +═══════════════════════════════════════════════════════════════════════════════ + [KREA 2] WEIGHT SLIDER +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Uncategorised + Author alcaitiff + Source https://civitai.com/models/2751675 + Provenance CivitAI · model 2751675 · version 3110226 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File WeightSlider-KREA2_v2.safetensors + Version v2.0 + Size 6.9 MB + Format SafeTensor / unknown precision + SHA256 8145f7315b5eb789bdca76665a5c1ace1f4f9883a1b41c7a02fc574be7234296 + Published 2026-07-08 + +─────────────────────────────────────────────────────────────────────────────── + 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 5,687 + Rating 100% positive (266/266) + Comments 14 + Published 58 days ago + Download rate 76.1/day since release + + Adoption rank #104 of 150 in this collection (Modest) + Momentum rank #36 of 150 by download rate (Trending) + + Adoption tier is the percentile of total downloads within this collection; + momentum is the percentile of downloads-per-day since release. Momentum is + ranked rather than measured against a fixed rate, because this base model + is itself new and every LoRA looks fast on an absolute scale. Both numbers + are CivitAI's, read on the scrape date at the foot of this file. + +─────────────────────────────────────────────────────────────────────────────── + CHAINING / STACKING +─────────────────────────────────────────────────────────────────────────────── + Stacks with others not stated + + The author gave no chaining guidance. The rules below are the + general Krea 2 ones, not anything specific to this LoRA. + + General Krea 2 rules that apply here (see REFERENCE.txt): + + • The most important rule is not about chaining but about which + checkpoint you chain onto: the official README says in bold "TRAIN on + Raw and RUN on Turbo", and "We highly recommend using RAW for training + LoRAs and applying them on Turbo for inference." LoRAs trained on Raw + are stated to work well on Turbo. + • Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 + LoRAs are DiT-only, so there is no CLIP side to chain and the text- + encoder branch stays untouched however many LoRAs you add. + • Krea 2 LoRAs stack in practice and the community treats multi-LoRA as + routine - style + character, style + slider, realism + character. + Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one + writes "I recently stacked over 20 of my own LoRAs on a single + generation and the output still came out great." + • Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge + on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is + loaded. The CivitAI 12GB training guide also notes rank affects + composability: "rank 8 produced a lighter style influence that was + easier to stack with other LoRAs, rank 16 provided a useful balance + between concept strength and composability, and rank 32 was better + suited to a precise object or concept intended to take priority over + other stacked LoRAs." + • Accelerator LoRAs are position-sensitive in intent, not in node order: + a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a + sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to- + Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is + explicitly recommended by the 4-step LoRA's author. + • Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI + (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card + says "Keep mu = 1.15. The training targets are anchored to that grid; a + different shift moves the [sampler off them]"; the same reasoning + applies to any Turbo-trained adapter in the chain. + • Reference/edit LoRAs are designed to be stacked under ordinary subject + and style LoRAs - the Identity Edit card states "Composes with your + LoRAs: character/body/style LoRAs stack on top and steer the prior" and + recommends a stacked subject LoRA when unusual subjects drift toward + the base prior. + • Krea 2 LoRAs cannot be chained with LoRAs for any other base model. + Many CivitAI pages carry multi-model titles like "(Krea-2 + ZIT)" or + "[Flux | ZIB | Krea2]" - those are separate files for separate + architectures shipped under one model page, not a combined adapter. + • To combine two full Krea 2 checkpoints rather than adapters, ComfyUI + ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per- + section weights for first., tmlp., txtmlp., tproj., + txtfusion.projector, txtfusion.layerwise_blocks.0-1, + txtfusion.refiner_blocks.0-1, blocks.0-27 and last. + • The author never states which weight family this targets. Silent + degradation, never a hard error, between Raw and Turbo. The two share + one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" + - so every LoRA tensor shape matches on both and ComfyUI/diffuser + +─────────────────────────────────────────────────────────────────────────────── + PROS +─────────────────────────────────────────────────────────────────────────────── + + Widely adopted — 5,687 downloads. [source: CivitAI stats] + + Positively received — 266 thumbs-up with no down-votes. [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] + - Thin documentation — expect trial and error. [source: description + length] + +─────────────────────────────────────────────────────────────────────────────── + COMPATIBILITY WARNING +─────────────────────────────────────────────────────────────────────────────── + Silent degradation, never a hard error, between Raw and Turbo. The two + share one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" - so + every LoRA tensor shape matches on both and ComfyUI/diffusers will load an + adapter against either without complaint. The lvladikov 4-step LoRA card + states this for the reverse direction: "Every layer it targets also exists + in Krea 2 Raw, so it will load there without complaint - but that is a side + effect of the shared architecture, not a supported mode", and reports + mixed, subject-dependent results with malformed subjects (duplicated heads, + fused limbs) at low step counts, needing 14+ steps with Raw's CFG on before + output is coherent. The intended direction (train on Raw, run on Turbo) is + officially blessed. The more common practical failure is a settings + mismatch rather than a weight mismatch: running a Turbo-targeted LoRA at + Raw's 52 steps + CFG 3.5, or leaving CFG enabled on Turbo, degrades output + with no warning. Loading a Krea 2 LoRA against a different model family + entirely (Flux, Qwen-Image, Z-Image) is also not an error - ComfyUI logs + "lora key not loaded: {key}" per key and generates as if no LoRA were + applied. + +─────────────────────────────────────────────────────────────────────────────── + AUTHOR'S DESCRIPTION +─────────────────────────────────────────────────────────────────────────────── + Use values from -3 to 3 but 2 is usually enough + +─────────────────────────────────────────────────────────────────────────────── + VERSION NOTES +─────────────────────────────────────────────────────────────────────────────── + Second version + +─────────────────────────────────────────────────────────────────────────────── + 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, 888x1184) + 02_image.jpeg (image, 1088x1448) + 03_image.jpeg (image, 888x1184) + 04_image.jpeg (image, 1440x1920) + 05_image.jpeg (image, 1440x1920) + 06_image.jpeg (image, 1440x1920) + +─────────────────────────────────────────────────────────────────────────────── + 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 00:18 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/krea-2-weight-slider/metadata.json b/krea-2-weight-slider/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..8ce5ddae499899d21a370d281a7a7ebe6ec30edf --- /dev/null +++ b/krea-2-weight-slider/metadata.json @@ -0,0 +1,134 @@ +{ + "name": "[KREA 2] Weight Slider", + "slug": "krea-2-weight-slider", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 2751675, + "version_id": 3110226, + "version_name": "v2.0", + "source": "civitai", + "url": "https://civitai.com/models/2751675", + "repo_id": null, + "file_path": null, + "author": "alcaitiff", + "category": { + "key": "other", + "label": "Uncategorised" + }, + "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 — 5,687 downloads. [source: CivitAI stats]", + "Positively received — 266 thumbs-up with no down-votes. [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]", + "Thin documentation — expect trial and error. [source: description length]" + ], + "file": { + "name": "WeightSlider-KREA2_v2.safetensors", + "size_bytes": 7202400, + "size_mb": 6.87, + "sha256": "8145f7315b5eb789bdca76665a5c1ace1f4f9883a1b41c7a02fc574be7234296" + }, + "stats": { + "downloads": 5687, + "thumbs_up": 266, + "thumbs_down": 0 + }, + "popularity": { + "downloads": 5687, + "thumbs_up": 266, + "thumbs_down": 0, + "comments": 14, + "approval": "100% positive (266/266)", + "age_days": 58.5, + "downloads_per_day": 76.14, + "rank_by_downloads": 104, + "rank_by_momentum": 36, + "of_total": 150, + "adoption_tier": "Modest", + "momentum": "Trending" + }, + "selection_score": 8879.0, + "selection_reasons": [ + "5,687 downloads", + "266 thumbs-up", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "888x1184", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "1088x1448", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "888x1184", + "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": "1440x1920", + "prompt": "" + } + ], + "tags": [ + "slider", + "tool", + "slim", + "fat", + "weight" + ], + "scraped_at": "2026-09-05T00:18:05+00:00", + "path": "krea-2/krea-2-weight-slider" +} diff --git a/meesh-style-krea-2-anima-illustrious/Meesh_Krea2_V01.safetensors b/meesh-style-krea-2-anima-illustrious/Meesh_Krea2_V01.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..25e4f3e4a88603646b045630113d7e0b3e0e8d05 --- /dev/null +++ b/meesh-style-krea-2-anima-illustrious/Meesh_Krea2_V01.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:63d3eb42ed319ebb37b16f64cc14aa777b536165519d06ab7cdaa68e5908d2ff +size 228591760 diff --git a/meesh-style-krea-2-anima-illustrious/example_images/01_image.jpeg b/meesh-style-krea-2-anima-illustrious/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..5985bac56a8f9ddb9d116683dcd92e1c933c8e66 --- /dev/null +++ b/meesh-style-krea-2-anima-illustrious/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:206f47c7ce9dbcd1db985a834b4f7308d19be8e0878d89fa936de9b680252648 +size 4438548 diff --git a/meesh-style-krea-2-anima-illustrious/example_images/02_image.jpeg b/meesh-style-krea-2-anima-illustrious/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..2f2220def945e45f23f73561294407b930a3e07c --- /dev/null +++ b/meesh-style-krea-2-anima-illustrious/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d141323ac66072b57617fafaa98e4ab2c2171015b8f73c46b75f1a6cf2d460f9 +size 2509602 diff --git a/meesh-style-krea-2-anima-illustrious/example_images/03_image.jpeg b/meesh-style-krea-2-anima-illustrious/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..dea92fda66ffb977279398c0e3d09ab89b82ba6d --- /dev/null +++ b/meesh-style-krea-2-anima-illustrious/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f50d48366cd328a9293b46dbed39a0a5027b6242ddc0c467cfc2cc4b1799a10d +size 2230974 diff --git a/meesh-style-krea-2-anima-illustrious/example_images/04_image.jpeg b/meesh-style-krea-2-anima-illustrious/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..f9b8dcae07f274799c386f15bb0ac0dc101f3338 --- /dev/null +++ b/meesh-style-krea-2-anima-illustrious/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f045d2d2fbbcf35e3a8f993eb6f56ff26c240e19288bb3b997aa9a1ea1b0ab20 +size 4175809 diff --git a/meesh-style-krea-2-anima-illustrious/example_images/05_image.jpeg b/meesh-style-krea-2-anima-illustrious/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..fc5a7df432a89d35517e0b8802f22e77d4a6886e --- /dev/null +++ b/meesh-style-krea-2-anima-illustrious/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8c462d0265ab2f7ab118a19312b86f40cc2a731cfa5a9f106e6df5634a36aea0 +size 933474 diff --git a/meesh-style-krea-2-anima-illustrious/example_images/06_image.jpeg b/meesh-style-krea-2-anima-illustrious/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..d6df9cf9aa90a2e7cad8f6ce18f459fd0d91103a --- /dev/null +++ b/meesh-style-krea-2-anima-illustrious/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ca1b252c61378173416c121ed620a7681c974bc8f54336db4b355ca9e84b58e +size 4209035 diff --git a/meesh-style-krea-2-anima-illustrious/info.txt b/meesh-style-krea-2-anima-illustrious/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..40144a0546585c68160022e965a020d9b35d4a7a --- /dev/null +++ b/meesh-style-krea-2-anima-illustrious/info.txt @@ -0,0 +1,296 @@ +═══════════════════════════════════════════════════════════════════════════════ + MEESH STYLE [KREA 2 / ANIMA / ILLUSTRIOUS] +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Visual style + Author Boerboer + Source https://civitai.com/models/2126663 + Provenance CivitAI · model 2126663 · version 3238418 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File Meesh_Krea2_V01.safetensors + Version v0.1 Krea 2 + Size 218.0 MB + Format SafeTensor / unknown precision + SHA256 63d3eb42ed319ebb37b16f64cc14aa777b536165519d06ab7cdaa68e5908d2ff + Published 2026-08-17 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + › "m33sh" + +─────────────────────────────────────────────────────────────────────────────── + 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 3,100 + Rating 100% positive (445/445) + Comments 10 + Published 19 days ago + Download rate 15.2/day since release + + Adoption rank #147 of 150 in this collection (Niche) + Momentum rank #116 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 +─────────────────────────────────────────────────────────────────────────────── + + Well established — 3,100 downloads. [source: CivitAI stats] + + Positively received — 445 thumbs-up with no down-votes. [source: + CivitAI stats] + + Explicit trigger word(s) — m33sh — 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 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 +─────────────────────────────────────────────────────────────────────────────── + This lora aims to replicate the style Meesh uses from 2017 onwards til + about early 2025 where his style switches again. + + The Krea version is a work in progress. I'm happy with it visually, but + it's prompt radiance could use some work. I'm happy enough with the Anima + version where I can call it a V1 and on par with the Illustrious version + (IMO). I will definitely keep an eye on ways to improve it further, but for + now, I hope you all enjoy! + +─────────────────────────────────────────────────────────────────────────────── + VERSION NOTES +─────────────────────────────────────────────────────────────────────────────── + Test version made with Krea 2. Use natural language for prompting. Also be + sure to use "anthropomorphic" instead of "anthro". Be sure to use the text + refusal reduction LoRA also if you want nsfw stuff. I've started + experimenting with porting some style LoRAs over to Krea 2. Don't let my + contest winner badge fool you, I don't really know what I'm doing yet. + Which is why I've labeled this as V0.1. For this version, I recommend using + natural language prompts. My next version will likely use tagging instead. + I'm going to continue trying new settings and techniques with other styles, + before hopefully revisiting this one for a proper V1. Enjoy in the + meanwhile! + +─────────────────────────────────────────────────────────────────────────────── + 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, 1632x2184) + 02_image.jpeg (image, 1344x1728) + 03_image.jpeg (image, 1085x1450) + 04_image.jpeg (image, 1632x2184) + 05_image.jpeg (image, 1152x768) + 06_image.jpeg (image, 1632x2184) + +─────────────────────────────────────────────────────────────────────────────── + 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 00:18 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/meesh-style-krea-2-anima-illustrious/metadata.json b/meesh-style-krea-2-anima-illustrious/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..3f514abe97914c56e084412200297742c99b5ccb --- /dev/null +++ b/meesh-style-krea-2-anima-illustrious/metadata.json @@ -0,0 +1,139 @@ +{ + "name": "Meesh Style [Krea 2 / Anima / Illustrious]", + "slug": "meesh-style-krea-2-anima-illustrious", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 2126663, + "version_id": 3238418, + "version_name": "v0.1 Krea 2", + "source": "civitai", + "url": "https://civitai.com/models/2126663", + "repo_id": null, + "file_path": null, + "author": "Boerboer", + "category": { + "key": "style", + "label": "Visual style" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [ + "m33sh" + ], + "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": [ + "Well established — 3,100 downloads. [source: CivitAI stats]", + "Positively received — 445 thumbs-up with no down-votes. [source: CivitAI stats]", + "Explicit trigger word(s) — m33sh — 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 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": "Meesh_Krea2_V01.safetensors", + "size_bytes": 228591760, + "size_mb": 218.0, + "sha256": "63d3eb42ed319ebb37b16f64cc14aa777b536165519d06ab7cdaa68e5908d2ff" + }, + "stats": { + "downloads": 3100, + "thumbs_up": 445, + "thumbs_down": 0 + }, + "popularity": { + "downloads": 3100, + "thumbs_up": 445, + "thumbs_down": 0, + "comments": 10, + "approval": "100% positive (445/445)", + "age_days": 18.5, + "downloads_per_day": 15.23, + "rank_by_downloads": 147, + "rank_by_momentum": 116, + "of_total": 150, + "adoption_tier": "Niche", + "momentum": "Long tail · fresh release" + }, + "selection_score": 8840.0, + "selection_reasons": [ + "3,100 downloads", + "445 thumbs-up", + "documented triggers + usage notes", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "1632x2184", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "1344x1728", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "1085x1450", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "1632x2184", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "1152x768", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "1632x2184", + "prompt": "" + } + ], + "tags": [ + "meesh", + "furry", + "male", + "bara", + "anthro", + "woman", + "style" + ], + "scraped_at": "2026-09-05T00:18:18+00:00", + "path": "krea-2/meesh-style-krea-2-anima-illustrious" +} diff --git a/mikkoani-style/example_images/01_image.jpeg b/mikkoani-style/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..9c310d88b983c8bca26c6144178ece3e26a30285 --- /dev/null +++ b/mikkoani-style/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd99e863c569ac6bc698a5237836396576dcaaa1d388a1c9f805d6c26e482150 +size 255467 diff --git a/mikkoani-style/example_images/02_image.jpeg b/mikkoani-style/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..f57de537511ef327fef7dac3ac599a6c6c9a8045 --- /dev/null +++ b/mikkoani-style/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9bb0d9a5749ef5268de4ce5a4ece4ed36aad3216967d0423fa878fa97f06341d +size 267981 diff --git a/mikkoani-style/example_images/03_image.jpeg b/mikkoani-style/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..a4f2d21afb2a65198f9d88ba734487e9a7dbfdc0 --- /dev/null +++ b/mikkoani-style/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:703412e6766e091f80f99ecb041e036ba7b5bd73ed14282acbf32438cfd935fa +size 245395 diff --git a/mikkoani-style/example_images/04_image.jpeg b/mikkoani-style/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..e1f8b985f4cd95d110f43ba6714265f2f680bf4d --- /dev/null +++ b/mikkoani-style/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d6a19330829a749134741ad85b4dc3270e8d89dce3a3d366b48891a2b82120ac +size 221489 diff --git a/mikkoani-style/example_images/05_image.jpeg b/mikkoani-style/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..f996a4dc7997fa87234bad4523cb4abc6d5299eb --- /dev/null +++ b/mikkoani-style/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4ade2393aee069acc6a72da8fbd3156e718aa2e50b3f11ba183f26157d8fdd63 +size 265555 diff --git a/mikkoani-style/example_images/06_image.jpeg b/mikkoani-style/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..0cf88d632b4f3a9e34c4a626e844d663180d020b --- /dev/null +++ b/mikkoani-style/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2be2d2439205d4a7651b57e9d2da3deb773248f6548495337b8e87c295b7effb +size 323303 diff --git a/mikkoani-style/info.txt b/mikkoani-style/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..51942673547c2fe38b44667d3edcd982f2bcb2a5 --- /dev/null +++ b/mikkoani-style/info.txt @@ -0,0 +1,276 @@ +═══════════════════════════════════════════════════════════════════════════════ + MIKKOANI STYLE +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode Text-to-image + Checkpoint Krea 2 RAW (krea/Krea-2-Raw) + Category Visual style + Author mikkoph + Source https://civitai.com/models/2470371 + Provenance CivitAI · model 2470371 · version 3065807 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File krea2-mikkoani.safetensors + Version Krea 2 v1.0 + Size 112.0 MB + Format SafeTensor / unknown precision + SHA256 3efb159513c1d2c1a2a6e022d51e8e168e191ec51972aed5d066a574cfc8078e + Published 2026-06-24 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + › "@mikkoani" + +─────────────────────────────────────────────────────────────────────────────── + RECOMMENDED SETTINGS +─────────────────────────────────────────────────────────────────────────────── + Taken verbatim from the author's description: + + Sampler steps 20–40 + CFG / guidance 4 + Sampler euler + +─────────────────────────────────────────────────────────────────────────────── + ADOPTION & POPULARITY +─────────────────────────────────────────────────────────────────────────────── + Downloads 4,845 + Rating 100% positive (314/314) + Comments 0 + Published 72 days ago + Download rate 12.6/day since release + + Adoption rank #119 of 150 in this collection (Niche) + Momentum rank #126 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. + • 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 +─────────────────────────────────────────────────────────────────────────────── + + Well established — 4,845 downloads. [source: CivitAI stats] + + Positively received — 314 thumbs-up with no down-votes. [source: + CivitAI stats] + + Explicit trigger word(s) — @mikkoani — so the effect can be turned on + and off from the prompt. [source: model metadata] + + Compact at 112 MB — cheap to stack with other LoRAs. [source: file + size] + + Author published concrete recommended settings. [source: description] + +─────────────────────────────────────────────────────────────────────────────── + CONS & CAVEATS +─────────────────────────────────────────────────────────────────────────────── + (none recorded) + +─────────────────────────────────────────────────────────────────────────────── + 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 +─────────────────────────────────────────────────────────────────────────────── + Anima + Can be used with base model with CFG=4-6, Steps 20-40 or using a Turbo LoRA + Works best with er_sde, euler, res_multistep and heun. + Most example images have "muted colors" and "complex linework" in the + prompt and no quality modifiers. + + Z-Image + Works with both Z-Image-Base + + https://civitai.com/models/2362961?modelVersionId=2741469 or with Z-Image + Turbo. + + The LoRA doesn't have a fixed trigger word (was trained with @mikkoani but + seems to be irrelevant) but can produce different styles depending on the + prompt. + + My favorite triggers are: + "anime illustration with muted color and complex linework" + "fashion sketch" + + With Z-Image-Base you get more freedom. With Z-Image-Turbo you'll need to + be a bit more convincing with your prompt to get the desired effect. + + Best used at 1.0, but can be pushed higher, especially with ZIT where at + least 1.5 is recommended. + + Klein 4B/9B + The LoRA works well both in editing and text-to-image mode. The 4b version + can be pushed a little stronger while the 9b version is already quite + strong at 1.0. It can give different variations depending on prompt, I like + to trigger it with + "@mikkoani illustration with muted colors and complex gritty linework". + Adding "anime" also works very 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, 832x1248) + 02_image.jpeg (image, 832x1248) + 03_image.jpeg (image, 832x1248) + 04_image.jpeg (image, 832x1248) + 05_image.jpeg (image, 1376x768) + 06_image.jpeg (image, 832x1248) + +─────────────────────────────────────────────────────────────────────────────── + 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 00:17 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/mikkoani-style/krea2-mikkoani.safetensors b/mikkoani-style/krea2-mikkoani.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..ae4ed4560ffc4866a5bf7d2aa34911458572b02a --- /dev/null +++ b/mikkoani-style/krea2-mikkoani.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3efb159513c1d2c1a2a6e022d51e8e168e191ec51972aed5d066a574cfc8078e +size 117395504 diff --git a/mikkoani-style/metadata.json b/mikkoani-style/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..8323e92efeb808402dc8fbcefc4f0afe7fc49167 --- /dev/null +++ b/mikkoani-style/metadata.json @@ -0,0 +1,143 @@ +{ + "name": "mikkoani Style", + "slug": "mikkoani-style", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 2470371, + "version_id": 3065807, + "version_name": "Krea 2 v1.0", + "source": "civitai", + "url": "https://civitai.com/models/2470371", + "repo_id": null, + "file_path": null, + "author": "mikkoph", + "category": { + "key": "style", + "label": "Visual style" + }, + "modalities": [ + "t2i" + ], + "primary_modality": "t2i", + "checkpoint_family": "raw", + "trigger_words": [ + "@mikkoani" + ], + "recommended_settings": { + "Sampler steps": "20–40", + "CFG / guidance": "4", + "Sampler": "euler" + }, + "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": [ + "Well established — 4,845 downloads. [source: CivitAI stats]", + "Positively received — 314 thumbs-up with no down-votes. [source: CivitAI stats]", + "Explicit trigger word(s) — @mikkoani — so the effect can be turned on and off from the prompt. [source: model metadata]", + "Compact at 112 MB — cheap to stack with other LoRAs. [source: file size]", + "Author published concrete recommended settings. [source: description]" + ], + "cons": [], + "file": { + "name": "krea2-mikkoani.safetensors", + "size_bytes": 117395504, + "size_mb": 111.96, + "sha256": "3efb159513c1d2c1a2a6e022d51e8e168e191ec51972aed5d066a574cfc8078e" + }, + "stats": { + "downloads": 4845, + "thumbs_up": 314, + "thumbs_down": 0 + }, + "popularity": { + "downloads": 4845, + "thumbs_up": 314, + "thumbs_down": 0, + "comments": 0, + "approval": "100% positive (314/314)", + "age_days": 72.4, + "downloads_per_day": 12.57, + "rank_by_downloads": 119, + "rank_by_momentum": 126, + "of_total": 150, + "adoption_tier": "Niche", + "momentum": "Long tail" + }, + "selection_score": 9013.0, + "selection_reasons": [ + "4,845 downloads", + "314 thumbs-up", + "documented triggers + usage notes", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "832x1248", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "832x1248", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "832x1248", + "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": "832x1248", + "prompt": "" + } + ], + "tags": [ + "style", + "horror illustration", + "complex linework", + "fashion sketch", + "anime", + "semirealistic", + "illustration" + ], + "scraped_at": "2026-09-05T00:17:59+00:00", + "path": "krea-2/mikkoani-style" +} diff --git a/rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b/RLY-thot_shot-KREA2-helena-v1-trigger-rlyhelena.safetensors b/rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b/RLY-thot_shot-KREA2-helena-v1-trigger-rlyhelena.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..067e043fa856d082b60e5e318d9f5de60859370d --- /dev/null +++ b/rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b/RLY-thot_shot-KREA2-helena-v1-trigger-rlyhelena.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:447f51db20ee54fa9eec88d052ed594324927ef79acd17226f107ee1b742ea99 +size 228587808 diff --git a/rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b/example_images/01_image.jpeg b/rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..75d7987656256b8d75db9b38a8b43fb97ae3151f --- /dev/null +++ b/rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5624d6f668d472d4b09bacf5994b73569291581a92e076e3a4fafd87134754cf +size 2193270 diff --git a/rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b/example_images/02_image.jpeg b/rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..6ca79277779b7fa78aa2b228e20379a0ad9cf57b --- /dev/null +++ b/rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e128c04258d2ab540d09dca075d75fb894027cc885bc523dcff4220fba26de5 +size 687632 diff --git a/rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b/example_images/03_image.jpeg b/rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..e26756fa498c61345efa30b2ce0cf0cab7691dc7 --- /dev/null +++ b/rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6af1175c4b6ad957d21e8c98e79ee6796d1cb4bbbaaa0e573b4ce219c8d10349 +size 18005770 diff --git a/rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b/example_images/04_poster.jpg b/rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b/example_images/04_poster.jpg new file mode 100644 index 0000000000000000000000000000000000000000..37c8e697670059f149ab4fcb1361952f9c6fe7ab Binary files /dev/null and b/rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b/example_images/04_poster.jpg differ diff --git a/rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b/example_images/04_video.mp4 b/rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b/example_images/04_video.mp4 new file mode 100644 index 0000000000000000000000000000000000000000..0d7b394b5ede83e238a0d584e41d31f98f1c4aab --- /dev/null +++ b/rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b/example_images/04_video.mp4 @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:23cee50df2a91917621b82f6aaa03c2bf509e05a0d3fed62b5a0b118e42fb60c +size 963588 diff --git a/rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b/example_images/05_image.jpeg b/rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..35240a1705fec093cffd835e1f06e9cf947c5d51 --- /dev/null +++ b/rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c74a9efcec6db6bc8d4a156f6d18380350e168e40c075f3d2106d6a768ca86cb +size 5589522 diff --git a/rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b/example_images/06_image.jpeg b/rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..d28997c35722898dfa3add248bc59fd9c94d26d0 --- /dev/null +++ b/rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2d627b6e93501cf3a93596bad35239222df223e66291a7153925560dae012e21 +size 399901 diff --git a/rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b/info.txt b/rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..72f935ec86c3ec858fb5acb73afeafdd41a3d355 --- /dev/null +++ b/rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b/info.txt @@ -0,0 +1,279 @@ +═══════════════════════════════════════════════════════════════════════════════ + RLY THOT SHOT - HELENA (KREA 2 + IDEOGRAM 4 + ZIB/ZIT + K9B) +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Photorealism + Author RaymondLuxuryYacht + Source https://civitai.com/models/2618083 + Provenance CivitAI · model 2618083 · version 3085180 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File RLY-thot_shot-KREA2-helena-v1-trigger-rlyhelena.safetensors + Version KREA2-v1.0 + Size 218.0 MB + Format SafeTensor / unknown precision + SHA256 447f51db20ee54fa9eec88d052ed594324927ef79acd17226f107ee1b742ea99 + Published 2026-06-30 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + › "rlyhelena" + +─────────────────────────────────────────────────────────────────────────────── + RECOMMENDED SETTINGS +─────────────────────────────────────────────────────────────────────────────── + Taken verbatim from the author's description: + + LoRA strength 1.0 + +─────────────────────────────────────────────────────────────────────────────── + ADOPTION & POPULARITY +─────────────────────────────────────────────────────────────────────────────── + Downloads 5,353 + Rating 100% positive (262/262) + Comments 10 + Published 66 days ago + Download rate 34.3/day since release + + Adoption rank #110 of 150 in this collection (Modest) + Momentum rank #70 of 150 by download rate (Steady) + + 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 — 5,353 downloads. [source: CivitAI stats] + + Positively received — 262 thumbs-up with no down-votes. [source: + CivitAI stats] + + Explicit trigger word(s) — rlyhelena — 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 labels this experimental — results vary shot to shot. [source: + description] + - Author references NSFW output — review before using in shared work. + [source: description] + - Author never states which generation mode it was trained for; test + against your own workflow before relying on it. [source: description] + +─────────────────────────────────────────────────────────────────────────────── + COMPATIBILITY WARNING +─────────────────────────────────────────────────────────────────────────────── + Silent degradation, never a hard error, between Raw and Turbo. The two + share one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" - so + every LoRA tensor shape matches on both and ComfyUI/diffusers will load an + adapter against either without complaint. The lvladikov 4-step LoRA card + states this for the reverse direction: "Every layer it targets also exists + in Krea 2 Raw, so it will load there without complaint - but that is a side + effect of the shared architecture, not a supported mode", and reports + mixed, subject-dependent results with malformed subjects (duplicated heads, + fused limbs) at low step counts, needing 14+ steps with Raw's CFG on before + output is coherent. The intended direction (train on Raw, run on Turbo) is + officially blessed. The more common practical failure is a settings + mismatch rather than a weight mismatch: running a Turbo-targeted LoRA at + Raw's 52 steps + CFG 3.5, or leaving CFG enabled on Turbo, degrades output + with no warning. Loading a Krea 2 LoRA against a different model family + entirely (Flux, Qwen-Image, Z-Image) is also not an error - ComfyUI logs + "lora key not loaded: {key}" per key and generates as if no LoRA were + applied. + +─────────────────────────────────────────────────────────────────────────────── + AUTHOR'S DESCRIPTION +─────────────────────────────────────────────────────────────────────────────── + Helena is the sixteenth release in a series ( RLY Thot Shots ) of female + character LoRAs. + Helena is a contradiction. From the Trailer Park in the American Deep South + to Milan, Italy. Raised in rural Alabama, now models for Vogue. Chews + tobacco and vapes but loves caviar and foie gras. Wears YSL to Mudfest. You + can take the girl out of the trailer park, but you can't take the trailer + park out of the girl. You'll catch her at the shooting range in heels. + Fun-fact: she's also Irena's half-sister (raised apart). + Getting started (basic) workflow here: + - RLY KREA 2 Starter Workflow + + - RLY Basic Starter ZiT Workflow + + Works best with its trained trigger: rlyhelena + Recommended strength: + - Z-Image: ~1.0 + + - Z-Image Turbo: 1.2+ . Feel free to experiment. + + - Ideogram 4: ~.7 (only load once, no unconditional) + + - KREA 2: 1.0 + + To share your own art and discuss all things (sfw and nsfw) realistic AI + image/music/video generation, join myself and other model creators and + enthusiasts on Unstable AI.sland: https://discord.gg/B6BSCbdAJX + +─────────────────────────────────────────────────────────────────────────────── + VERSION NOTES +─────────────────────────────────────────────────────────────────────────────── + Trained on KREA 2 (RAW). Use with Turbo model at Strength 1.0. + +─────────────────────────────────────────────────────────────────────────────── + 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, 1126x1644) + 02_image.jpeg (image, 1776x2368) + 03_image.jpeg (image, 3328x4864) + 04_video.mp4 (video, 544x960) + still: 04_poster.jpg + 05_image.jpeg (image, 1328x1944) + 06_image.jpeg (image, 1352x1824) + +─────────────────────────────────────────────────────────────────────────────── + 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 00:18 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b/metadata.json b/rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..e1066b854d6a426ff98979ba3b85d6f6ee6007cf --- /dev/null +++ b/rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b/metadata.json @@ -0,0 +1,142 @@ +{ + "name": "RLY Thot Shot - Helena (KREA 2 + Ideogram 4 + ZiB/ZiT + K9B)", + "slug": "rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 2618083, + "version_id": 3085180, + "version_name": "KREA2-v1.0", + "source": "civitai", + "url": "https://civitai.com/models/2618083", + "repo_id": null, + "file_path": null, + "author": "RaymondLuxuryYacht", + "category": { + "key": "photoreal", + "label": "Photorealism" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [ + "rlyhelena" + ], + "recommended_settings": { + "LoRA strength": "1.0" + }, + "chaining": { + "author": { + "stacks": "not stated", + "notes": "", + "avoid": [], + "requires": [], + "order": "", + "strength": "", + "stated": false + }, + "general_rules": [ + "The most important rule is not about chaining but about which checkpoint you chain onto: the official README says in bold \"TRAIN on Raw and RUN on Turbo\", and \"We highly recommend using RAW for training LoRAs and applying them on Turbo for inference.\" LoRAs trained on Raw are stated to work well on Turbo.", + "Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 LoRAs are DiT-only, so there is no CLIP side to chain and the text-encoder branch stays untouched however many LoRAs you add.", + "Krea 2 LoRAs stack in practice and the community treats multi-LoRA as routine - style + character, style + slider, realism + character. Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one writes \"I recently stacked over 20 of my own LoRAs on a single generation and the output still came out great.\"", + "Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is loaded. The CivitAI 12GB training guide also notes rank affects composability: \"rank 8 produced a lighter style influence that was easier to stack with other LoRAs, rank 16 provided a useful balance between concept strength and composability, and rank 32 was better suited to a precise object or concept intended to take priority over other stacked LoRAs.\"", + "Accelerator LoRAs are position-sensitive in intent, not in node order: a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to-Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is explicitly recommended by the 4-step LoRA's author.", + "Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card says \"Keep mu = 1.15. The training targets are anchored to that grid; a different shift moves the [sampler off them]\"; the same reasoning applies to any Turbo-trained adapter in the chain.", + "Reference/edit LoRAs are designed to be stacked under ordinary subject and style LoRAs - the Identity Edit card states \"Composes with your LoRAs: character/body/style LoRAs stack on top and steer the prior\" and recommends a stacked subject LoRA when unusual subjects drift toward the base prior.", + "Krea 2 LoRAs cannot be chained with LoRAs for any other base model. Many CivitAI pages carry multi-model titles like \"(Krea-2 + ZIT)\" or \"[Flux | ZIB | Krea2]\" - those are separate files for separate architectures shipped under one model page, not a combined adapter.", + "To combine two full Krea 2 checkpoints rather than adapters, ComfyUI ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per-section weights for first., tmlp., txtmlp., tproj., txtfusion.projector, txtfusion.layerwise_blocks.0-1, txtfusion.refiner_blocks.0-1, blocks.0-27 and last.", + "The author never states which weight family this targets. Silent degradation, never a hard error, between Raw and Turbo. The two share one architecture - ai-toolkit's config is commented \"The reference 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)\" - so every LoRA tensor shape matches on both and ComfyUI/diffuser" + ] + }, + "pros": [ + "Widely adopted — 5,353 downloads. [source: CivitAI stats]", + "Positively received — 262 thumbs-up with no down-votes. [source: CivitAI stats]", + "Explicit trigger word(s) — rlyhelena — 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 labels this experimental — results vary shot to shot. [source: description]", + "Author references NSFW output — review before using in shared work. [source: description]", + "Author never states which generation mode it was trained for; test against your own workflow before relying on it. [source: description]" + ], + "file": { + "name": "RLY-thot_shot-KREA2-helena-v1-trigger-rlyhelena.safetensors", + "size_bytes": 228587808, + "size_mb": 218.0, + "sha256": "447f51db20ee54fa9eec88d052ed594324927ef79acd17226f107ee1b742ea99" + }, + "stats": { + "downloads": 5353, + "thumbs_up": 262, + "thumbs_down": 0 + }, + "popularity": { + "downloads": 5353, + "thumbs_up": 262, + "thumbs_down": 0, + "comments": 10, + "approval": "100% positive (262/262)", + "age_days": 66.3, + "downloads_per_day": 34.26, + "rank_by_downloads": 110, + "rank_by_momentum": 70, + "of_total": 150, + "adoption_tier": "Modest", + "momentum": "Steady" + }, + "selection_score": 8897.0, + "selection_reasons": [ + "5,353 downloads", + "262 thumbs-up", + "documented triggers + usage notes", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "1126x1644", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "1776x2368", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "3328x4864", + "prompt": "" + }, + { + "file": "04_video.mp4", + "type": "video", + "resolution": "544x960", + "prompt": "", + "poster": "04_poster.jpg" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "1328x1944", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "1352x1824", + "prompt": "" + } + ], + "tags": [ + "woman", + "character", + "rly", + "photorealistic", + "photography" + ], + "scraped_at": "2026-09-05T00:18:09+00:00", + "path": "krea-2/rly-thot-shot-helena-krea-2-ideogram-4-zib-zit-k9b" +}