diff --git a/.gitattributes b/.gitattributes index f0d151691ec715e97a5af83e8eea7d89c561339c..4dcadbf4e6b6d3dd49e21cb9bc4a217adc8ddb17 100644 --- a/.gitattributes +++ b/.gitattributes @@ -525,3 +525,75 @@ z-female-body-retouched/example_images/01_image.jpeg filter=lfs diff=lfs merge=l z-female-body-retouched/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text z-female-body-retouched/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text z-female-body-retouched/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +glowing-and-light-particles/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +glowing-and-light-particles/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +glowing-and-light-particles/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +glowing-and-light-particles/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +glowing-and-light-particles/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +glowing-and-light-particles/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +purelens-realism-ideogram-4-krea-2/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +purelens-realism-ideogram-4-krea-2/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +purelens-realism-ideogram-4-krea-2/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +purelens-realism-ideogram-4-krea-2/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +purelens-realism-ideogram-4-krea-2/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +purelens-realism-ideogram-4-krea-2/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +minimalist-vector-art/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +minimalist-vector-art/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +minimalist-vector-art/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +minimalist-vector-art/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +minimalist-vector-art/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +minimalist-vector-art/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +da-vinci-drawing-style-ce/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +da-vinci-drawing-style-ce/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +da-vinci-drawing-style-ce/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +da-vinci-drawing-style-ce/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +da-vinci-drawing-style-ce/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +da-vinci-drawing-style-ce/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +yfg-chatgpt-4o-style-flux-zit-krea2/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +yfg-chatgpt-4o-style-flux-zit-krea2/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +yfg-chatgpt-4o-style-flux-zit-krea2/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +yfg-chatgpt-4o-style-flux-zit-krea2/example_images/04_video.mp4 filter=lfs diff=lfs merge=lfs -text +yfg-chatgpt-4o-style-flux-zit-krea2/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +yfg-chatgpt-4o-style-flux-zit-krea2/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +age-slider/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +age-slider/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +age-slider/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +age-slider/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +age-slider/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +age-slider/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +krea-2-realism-slider/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +krea-2-realism-slider/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +krea-2-realism-slider/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +krea-2-realism-slider/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +krea-2-realism-slider/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +krea-2-realism-slider/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +elegance-niji/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +elegance-niji/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +elegance-niji/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +elegance-niji/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +elegance-niji/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +elegance-niji/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +manga-style-naoki-urasawa/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +manga-style-naoki-urasawa/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +manga-style-naoki-urasawa/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +manga-style-naoki-urasawa/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +manga-style-naoki-urasawa/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +manga-style-naoki-urasawa/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +krea2-qwen2511-z-image-illustria-anime-01/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +krea2-qwen2511-z-image-illustria-anime-01/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +krea2-qwen2511-z-image-illustria-anime-01/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +krea2-qwen2511-z-image-illustria-anime-01/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +krea2-qwen2511-z-image-illustria-anime-01/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +krea2-qwen2511-z-image-illustria-anime-01/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +ray-krea2-style-loras/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +ray-krea2-style-loras/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +ray-krea2-style-loras/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +ray-krea2-style-loras/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +ray-krea2-style-loras/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +ray-krea2-style-loras/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text +harness-panel-gag/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text +harness-panel-gag/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text +harness-panel-gag/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text +harness-panel-gag/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text +harness-panel-gag/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text +harness-panel-gag/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text diff --git a/age-slider/Age_Slider_krea2t_000000020.safetensors b/age-slider/Age_Slider_krea2t_000000020.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..42cb90effa5c3bc9a3c426aa62e27dea5a754f0c --- /dev/null +++ b/age-slider/Age_Slider_krea2t_000000020.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:01940230df88ef880e0e24cd6a8ba6890a81cc81e0227d25f1c950aa0866e6b4 +size 28627752 diff --git a/age-slider/example_images/01_image.jpeg b/age-slider/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..337c2144e9e98aaf13a91fd724ca5e66ea4f8add --- /dev/null +++ b/age-slider/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2be71bb31067dd4afc097fce8c90753ef61aab5c39c9fa0448316acc0a86fe3a +size 2464431 diff --git a/age-slider/example_images/02_image.jpeg b/age-slider/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..cfab26712e42312c4553a7b97aee590d38294b27 --- /dev/null +++ b/age-slider/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:684a17b72643912a229c5293db8fc40c7636acc979f753365e0232857e871e16 +size 4354311 diff --git a/age-slider/example_images/03_image.jpeg b/age-slider/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..def76434f5897845d457a5f5ea4203e2890e44d1 --- /dev/null +++ b/age-slider/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c6dc8a9e4736459d9a0c51381b23030865d4b72fc3e61ba293e2a0e26176ae94 +size 4181485 diff --git a/age-slider/example_images/04_image.jpeg b/age-slider/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..f06695ae830ce2afb7c77472f8e1e60e1d7aaf88 --- /dev/null +++ b/age-slider/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2cb4c5774162c054a5a94021a6d6fb8d16e8cd4e533abddbdf90d7a8c996ee70 +size 5535369 diff --git a/age-slider/example_images/05_image.jpeg b/age-slider/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..2608d1056b2b75ae325f1a40f2690762d0108356 --- /dev/null +++ b/age-slider/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b3440e3fc2765b0f906e7fdef35a2f6edb6f44bf3e833a881116bd6c0b9fc09f +size 3220817 diff --git a/age-slider/example_images/06_image.jpeg b/age-slider/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..0a58291dae1d9fbefb9b671743c2191197f76754 --- /dev/null +++ b/age-slider/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:49fd42534bf99705c8e56c619430ed43d8634ec5fa0dea472422b7c3a69c3457 +size 2363284 diff --git a/age-slider/info.txt b/age-slider/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..272eab65532e3c195e101064f34534ebfbc3b63a --- /dev/null +++ b/age-slider/info.txt @@ -0,0 +1,280 @@ +═══════════════════════════════════════════════════════════════════════════════ + AGE SLIDER +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Uncategorised + Author el_chupanibre + Source https://civitai.com/models/2633687 + Provenance CivitAI · model 2633687 · version 3139661 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File Age_Slider_krea2t_000000020.safetensors + Version Krea2 + Size 27.3 MB + Format SafeTensor / unknown precision + SHA256 01940230df88ef880e0e24cd6a8ba6890a81cc81e0227d25f1c950aa0866e6b4 + Published 2026-07-17 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + No trigger word. The LoRA applies as soon as it is loaded — + strength is the only way to dial the effect up or down. + +─────────────────────────────────────────────────────────────────────────────── + RECOMMENDED SETTINGS +─────────────────────────────────────────────────────────────────────────────── + The author did not publish specific settings. + No author settings were published. Community baseline: steps Turbo: 8 + (official README and HF card CLI example; the ComfyUI template ships 8). + Raw: 52 (official README). The krea-2 CLI's generic default is 28. + Community Turbo LoRAs commonly suggest 8-12 (e.g. the Identity Edit card + recommends Turbo at 8-12, or Raw at 20 for removals). A community 4-step + LoRA on Turbo makes 4 valid; its author calls 2 steps an out-of-spec + preview mode., cfg Turbo: CFG disabled - guidance_scale 0.0 in diffusers + and --cfg 0.0 in the krea-2 CLI, which is cfg 1.0 in ComfyUI's KSampler + (the official template ships cfg 1). These are the same thing; do not set + ComfyUI cfg to 0. Raw: CFG 3.5 (official README; CLI generic default 4.5). + Negative prompting on Turbo is effectively unused - the official templates + route ConditioningZeroOut into the negative slot., sampler ComfyUI: + KSampler with sampler euler, scheduler simple, denoise 1.0 (verified in + image_krea2_turbo_t2i.json: [seed, 'randomize', 8, 1, 'euler', 'simple', + 1]). Timestep shift mu = 1.15 for Turbo; ComfyUI bakes this in via + Krea2.sampling_settings = {"multiplier": 1.0, "shift": 1.15}, so the plain + t2i template needs no ModelSampling node, while the style-reference + template adds ModelSamplingFlux at max_shift 1.15 / base_shift 0.5. The CLI + exposes --mu 1.15 for Turbo; Raw uses a resolution-derived mu interpolated + between --y1 0.5 and --y2 1.15 (ai-toolkit mirrors this as base_shift 0.5 / + max_shift 1.15 over image_seq_len 256-6400, exponential time shift)., + resolution 1024x1024 default. Turbo: 1K-2K (the official CLI example + generates 2048x2048; ComfyUI's ResolutionSelector defaults to 1 megapixel, + set 2.0 for 2K). Raw: "trained to generate upto 1k resolution". + Width/height are padded up to a multiple of 16., lora_strength 1.0 for + Krea's nine official style LoRAs (docs.comfy.org lists Recommended Strength + 1.0 for every one), though the shipped t2i template's LoraLoaderModelOnly + is set to 0.8. The style-reference LoRA is used at 1.0 in its template. The + 4-step distill LoRA documents 1.0 as its trained point, 1.0-1.5 as coherent + extrapolation and above ~1.5 as speckle artefacting. Community subject + LoRAs are typically recommended around 0.8-1.2 and dropped toward 0.6-0.9 + when several are stacked.. + +─────────────────────────────────────────────────────────────────────────────── + ADOPTION & POPULARITY +─────────────────────────────────────────────────────────────────────────────── + Downloads 9,166 + Rating 100% positive (232/232) + Comments 0 + Published 49 days ago + Download rate 89.0/day since release + + Adoption rank #61 of 150 in this collection (Moderate) + Momentum rank #31 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 — 9,166 downloads. [source: CivitAI stats] + + Positively received — 232 thumbs-up with no down-votes. [source: + CivitAI stats] + + Compact at 27 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 +─────────────────────────────────────────────────────────────────────────────── + Controls the age. + + (-) Younger + (+) Older + + except for the Flux.1 Version B it's vice versa and female only + +─────────────────────────────────────────────────────────────────────────────── + 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, 1920x2176) + 02_image.jpeg (image, 1472x2048) + 03_image.jpeg (image, 1472x2048) + 04_image.jpeg (image, 1472x2048) + 05_image.jpeg (image, 1200x1800) + 06_image.jpeg (image, 1200x1696) + +─────────────────────────────────────────────────────────────────────────────── + 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 00:16 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/age-slider/metadata.json b/age-slider/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..d63323f6a77de5de17a2674822a030618b230578 --- /dev/null +++ b/age-slider/metadata.json @@ -0,0 +1,131 @@ +{ + "name": "Age Slider", + "slug": "age-slider", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 2633687, + "version_id": 3139661, + "version_name": "Krea2", + "source": "civitai", + "url": "https://civitai.com/models/2633687", + "repo_id": null, + "file_path": null, + "author": "el_chupanibre", + "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 — 9,166 downloads. [source: CivitAI stats]", + "Positively received — 232 thumbs-up with no down-votes. [source: CivitAI stats]", + "Compact at 27 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": "Age_Slider_krea2t_000000020.safetensors", + "size_bytes": 28627752, + "size_mb": 27.3, + "sha256": "01940230df88ef880e0e24cd6a8ba6890a81cc81e0227d25f1c950aa0866e6b4" + }, + "stats": { + "downloads": 9166, + "thumbs_up": 232, + "thumbs_down": 0 + }, + "popularity": { + "downloads": 9166, + "thumbs_up": 232, + "thumbs_down": 0, + "comments": 0, + "approval": "100% positive (232/232)", + "age_days": 49.1, + "downloads_per_day": 88.97, + "rank_by_downloads": 61, + "rank_by_momentum": 31, + "of_total": 150, + "adoption_tier": "Moderate", + "momentum": "Trending" + }, + "selection_score": 11950.0, + "selection_reasons": [ + "9,166 downloads", + "232 thumbs-up", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "1920x2176", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "1472x2048", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "1472x2048", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "1472x2048", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "1200x1800", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "1200x1696", + "prompt": "" + } + ], + "tags": [ + "age slider", + "concept" + ], + "scraped_at": "2026-09-05T00:16:01+00:00", + "path": "krea-2/age-slider" +} diff --git a/da-vinci-drawing-style-ce/DaVinciDrawing01_CE_Krea2_AIT5k.safetensors b/da-vinci-drawing-style-ce/DaVinciDrawing01_CE_Krea2_AIT5k.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..fc82e804af0e84f4b60d5424293a11be170a02bb --- /dev/null +++ b/da-vinci-drawing-style-ce/DaVinciDrawing01_CE_Krea2_AIT5k.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:56f8862c81a767836f395c1d9f16a6339f19632c7bf9663b37aacdb066ccb264 +size 228587824 diff --git a/da-vinci-drawing-style-ce/example_images/01_image.jpeg b/da-vinci-drawing-style-ce/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..3d4b49398a86dc9cd9aff57b24357260ba8cd6d3 --- /dev/null +++ b/da-vinci-drawing-style-ce/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7661b63a10f0f4c7e0c086d2f5f9c63a519096fc409c92e7b2ac3d9d87ddf430 +size 1934518 diff --git a/da-vinci-drawing-style-ce/example_images/02_image.jpeg b/da-vinci-drawing-style-ce/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..6685b105af8d86fbaa766203be3ac2d9d238ab45 --- /dev/null +++ b/da-vinci-drawing-style-ce/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dcb23bd1aea3fd37a301e729d37f8b9845bc22ff1eeaecb3bceac1d78902cc59 +size 3079198 diff --git a/da-vinci-drawing-style-ce/example_images/03_image.jpeg b/da-vinci-drawing-style-ce/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..1e775ec7308b3b1f41aed8d58319eba1e8ad33ce --- /dev/null +++ b/da-vinci-drawing-style-ce/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb7f8d0c0378423df45d5c9a605b4433e9803dfc42db920f6f8fda90295448c1 +size 3072958 diff --git a/da-vinci-drawing-style-ce/example_images/04_image.jpeg b/da-vinci-drawing-style-ce/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..2dcf7e17db09f1b4bd4be38eb8baf82f29a82fb3 --- /dev/null +++ b/da-vinci-drawing-style-ce/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:85408b1f6c9b9c8f248cfc4e7f9f28a1d4a5a322cc844c0bd44ea1ed2b2640b2 +size 3092952 diff --git a/da-vinci-drawing-style-ce/example_images/05_image.jpeg b/da-vinci-drawing-style-ce/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..22e0fa8b759b92c39c58c8e353dade6076f558c2 --- /dev/null +++ b/da-vinci-drawing-style-ce/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f20097264fd83d58ecece6e086a84958936acb57054c2c151050635e601d6260 +size 3043968 diff --git a/da-vinci-drawing-style-ce/example_images/06_image.jpeg b/da-vinci-drawing-style-ce/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..e27e24e80bc7a795c66cd0fcdb8f76b6833184dc --- /dev/null +++ b/da-vinci-drawing-style-ce/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:87221fe01844951521e43b38685ff552a154f6759f6aa2fa44a105d2477aa793 +size 3087625 diff --git a/da-vinci-drawing-style-ce/info.txt b/da-vinci-drawing-style-ce/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..f4bb5699e9e25528c2f233d0e051c6d799227e77 --- /dev/null +++ b/da-vinci-drawing-style-ce/info.txt @@ -0,0 +1,254 @@ +═══════════════════════════════════════════════════════════════════════════════ + DA VINCI DRAWING STYLE - CE +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Visual style + Author CreativeEdge + Source https://civitai.com/models/671809 + Provenance CivitAI · model 671809 · version 3229828 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File DaVinciDrawing01_CE_Krea2_AIT5k.safetensors + Version V01 - Krea2 Raw/Turbo + Size 218.0 MB + Format SafeTensor / unknown precision + SHA256 56f8862c81a767836f395c1d9f16a6339f19632c7bf9663b37aacdb066ccb264 + Published 2026-08-15 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + › "dvnciCE_style" + +─────────────────────────────────────────────────────────────────────────────── + RECOMMENDED SETTINGS +─────────────────────────────────────────────────────────────────────────────── + Taken verbatim from the author's description: + + LoRA strength 0.9 + +─────────────────────────────────────────────────────────────────────────────── + ADOPTION & POPULARITY +─────────────────────────────────────────────────────────────────────────────── + Downloads 6,288 + Rating 100% positive (501/502) + Comments 3 + Published 20 days ago + Download rate 13.4/day since release + + Adoption rank #90 of 150 in this collection (Modest) + Momentum rank #121 of 150 by download rate (Long tail · fresh release) + + Adoption tier is the percentile of total downloads within this collection; + momentum is the percentile of downloads-per-day since release. Momentum is + ranked rather than measured against a fixed rate, because this base model + is itself new and every LoRA looks fast on an absolute scale. Both numbers + are CivitAI's, read on the scrape date at the foot of this file. + +─────────────────────────────────────────────────────────────────────────────── + CHAINING / STACKING +─────────────────────────────────────────────────────────────────────────────── + Stacks with others not stated + + The author gave no chaining guidance. The rules below are the + general Krea 2 ones, not anything specific to this LoRA. + + General Krea 2 rules that apply here (see REFERENCE.txt): + + • The most important rule is not about chaining but about which + checkpoint you chain onto: the official README says in bold "TRAIN on + Raw and RUN on Turbo", and "We highly recommend using RAW for training + LoRAs and applying them on Turbo for inference." LoRAs trained on Raw + are stated to work well on Turbo. + • Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 + LoRAs are DiT-only, so there is no CLIP side to chain and the text- + encoder branch stays untouched however many LoRAs you add. + • Krea 2 LoRAs stack in practice and the community treats multi-LoRA as + routine - style + character, style + slider, realism + character. + Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one + writes "I recently stacked over 20 of my own LoRAs on a single + generation and the output still came out great." + • Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge + on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is + loaded. The CivitAI 12GB training guide also notes rank affects + composability: "rank 8 produced a lighter style influence that was + easier to stack with other LoRAs, rank 16 provided a useful balance + between concept strength and composability, and rank 32 was better + suited to a precise object or concept intended to take priority over + other stacked LoRAs." + • Accelerator LoRAs are position-sensitive in intent, not in node order: + a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a + sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to- + Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is + explicitly recommended by the 4-step LoRA's author. + • Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI + (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card + says "Keep mu = 1.15. The training targets are anchored to that grid; a + different shift moves the [sampler off them]"; the same reasoning + applies to any Turbo-trained adapter in the chain. + • Reference/edit LoRAs are designed to be stacked under ordinary subject + and style LoRAs - the Identity Edit card states "Composes with your + LoRAs: character/body/style LoRAs stack on top and steer the prior" and + recommends a stacked subject LoRA when unusual subjects drift toward + the base prior. + • Krea 2 LoRAs cannot be chained with LoRAs for any other base model. + Many CivitAI pages carry multi-model titles like "(Krea-2 + ZIT)" or + "[Flux | ZIB | Krea2]" - those are separate files for separate + architectures shipped under one model page, not a combined adapter. + • To combine two full Krea 2 checkpoints rather than adapters, ComfyUI + ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per- + section weights for first., tmlp., txtmlp., tproj., + txtfusion.projector, txtfusion.layerwise_blocks.0-1, + txtfusion.refiner_blocks.0-1, blocks.0-27 and last. + • The author never states which weight family this targets. Silent + degradation, never a hard error, between Raw and Turbo. The two share + one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" + - so every LoRA tensor shape matches on both and ComfyUI/diffuser + +─────────────────────────────────────────────────────────────────────────────── + PROS +─────────────────────────────────────────────────────────────────────────────── + + Widely adopted — 6,288 downloads. [source: CivitAI stats] + + Positively received — 501 thumbs-up, 1 down. [source: CivitAI stats] + + Explicit trigger word(s) — dvnciCE_style — so the effect can be turned + on and off from the prompt. [source: model metadata] + + Compact at 218 MB — cheap to stack with other LoRAs. [source: file + size] + + Author published concrete recommended settings. [source: description] + +─────────────────────────────────────────────────────────────────────────────── + CONS & CAVEATS +─────────────────────────────────────────────────────────────────────────────── + - Author never states which generation mode it was trained for; test + against your own workflow before relying on it. [source: description] + +─────────────────────────────────────────────────────────────────────────────── + COMPATIBILITY WARNING +─────────────────────────────────────────────────────────────────────────────── + Silent degradation, never a hard error, between Raw and Turbo. The two + share one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" - so + every LoRA tensor shape matches on both and ComfyUI/diffusers will load an + adapter against either without complaint. The lvladikov 4-step LoRA card + states this for the reverse direction: "Every layer it targets also exists + in Krea 2 Raw, so it will load there without complaint - but that is a side + effect of the shared architecture, not a supported mode", and reports + mixed, subject-dependent results with malformed subjects (duplicated heads, + fused limbs) at low step counts, needing 14+ steps with Raw's CFG on before + output is coherent. The intended direction (train on Raw, run on Turbo) is + officially blessed. The more common practical failure is a settings + mismatch rather than a weight mismatch: running a Turbo-targeted LoRA at + Raw's 52 steps + CFG 3.5, or leaving CFG enabled on Turbo, degrades output + with no warning. Loading a Krea 2 LoRA against a different model family + entirely (Flux, Qwen-Image, Z-Image) is also not an error - ComfyUI logs + "lora key not loaded: {key}" per key and generates as if no LoRA were + applied. + +─────────────────────────────────────────────────────────────────────────────── + AUTHOR'S DESCRIPTION +─────────────────────────────────────────────────────────────────────────────── + Please share your creations. I would love to see how you have used the + LoRA. + Weight around 0.9 suits most pictures, within a range of around 0.5 - 1.2 + depending on desired result. + +─────────────────────────────────────────────────────────────────────────────── + VERSION NOTES +─────────────────────────────────────────────────────────────────────────────── + The LoRA is created using the raw version. It is suitable for both raw and + turbo versions. All showcase images were generated using the raw model. + Note: The LoRA has a trigger word. However, the trigger word often appears + as text in the resultant image, so I advise against using it. + +─────────────────────────────────────────────────────────────────────────────── + EXAMPLE MEDIA +─────────────────────────────────────────────────────────────────────────────── + 6 preview(s) in ./example_images/ — community results for + this LoRA. Videos keep a matching _poster.jpg still frame. + + No prompts are recorded below: CivitAI's public API flags these posts as + having generation metadata but does not return it, so the prompts cannot be + scraped. Open the model page to read them. + + 01_image.jpeg (image, 840x1256) + 02_image.jpeg (image, 1216x1216) + 03_image.jpeg (image, 1216x1216) + 04_image.jpeg (image, 1216x1216) + 05_image.jpeg (image, 1216x1216) + 06_image.jpeg (image, 1216x1216) + +─────────────────────────────────────────────────────────────────────────────── + 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 00:15 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/da-vinci-drawing-style-ce/metadata.json b/da-vinci-drawing-style-ce/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..9a621420956482a0130148c869896f95c9329d57 --- /dev/null +++ b/da-vinci-drawing-style-ce/metadata.json @@ -0,0 +1,136 @@ +{ + "name": "Da Vinci Drawing style - CE", + "slug": "da-vinci-drawing-style-ce", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 671809, + "version_id": 3229828, + "version_name": "V01 - Krea2 Raw/Turbo", + "source": "civitai", + "url": "https://civitai.com/models/671809", + "repo_id": null, + "file_path": null, + "author": "CreativeEdge", + "category": { + "key": "style", + "label": "Visual style" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [ + "dvnciCE_style" + ], + "recommended_settings": { + "LoRA strength": "0.9" + }, + "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 — 6,288 downloads. [source: CivitAI stats]", + "Positively received — 501 thumbs-up, 1 down. [source: CivitAI stats]", + "Explicit trigger word(s) — dvnciCE_style — so the effect can be turned on and off from the prompt. [source: model metadata]", + "Compact at 218 MB — cheap to stack with other LoRAs. [source: file size]", + "Author published concrete recommended settings. [source: description]" + ], + "cons": [ + "Author never states which generation mode it was trained for; test against your own workflow before relying on it. [source: description]" + ], + "file": { + "name": "DaVinciDrawing01_CE_Krea2_AIT5k.safetensors", + "size_bytes": 228587824, + "size_mb": 218.0, + "sha256": "56f8862c81a767836f395c1d9f16a6339f19632c7bf9663b37aacdb066ccb264" + }, + "stats": { + "downloads": 6288, + "thumbs_up": 501, + "thumbs_down": 1 + }, + "popularity": { + "downloads": 6288, + "thumbs_up": 501, + "thumbs_down": 1, + "comments": 3, + "approval": "100% positive (501/502)", + "age_days": 20.0, + "downloads_per_day": 13.39, + "rank_by_downloads": 90, + "rank_by_momentum": 121, + "of_total": 150, + "adoption_tier": "Modest", + "momentum": "Long tail · fresh release" + }, + "selection_score": 12300.0, + "selection_reasons": [ + "6,288 downloads", + "501 thumbs-up", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "840x1256", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "1216x1216", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "1216x1216", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "1216x1216", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "1216x1216", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "1216x1216", + "prompt": "" + } + ], + "tags": [ + "drawing", + "style", + "da vinci" + ], + "scraped_at": "2026-09-05T00:15:55+00:00", + "path": "krea-2/da-vinci-drawing-style-ce" +} diff --git a/elegance-niji/Elegance_Niji_KR2.safetensors b/elegance-niji/Elegance_Niji_KR2.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..8282cbbda9e2504e24d6b5e464b8ac6d36883757 --- /dev/null +++ b/elegance-niji/Elegance_Niji_KR2.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a71faadb8a68220eb28883ee7ef7d0317a63b0995ce45a76fe927cd99675ceb +size 228588856 diff --git a/elegance-niji/example_images/01_image.jpeg b/elegance-niji/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..faaaef6f42ac6bfd982614457c696fcad7ecfe38 --- /dev/null +++ b/elegance-niji/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:368ed8a1fd71d6fb524c4b1c463ee11ce88818860b7fd5e62200052a94df21ed +size 3209763 diff --git a/elegance-niji/example_images/02_image.jpeg b/elegance-niji/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..7e466e79e760517814bcb3e417da54366f09e5fd --- /dev/null +++ b/elegance-niji/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7292eddb68bff3a338af9010cf6e885d37995cb8c6b0d3356aa10abdaf13c674 +size 11317442 diff --git a/elegance-niji/example_images/03_image.jpeg b/elegance-niji/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..12b81ee61a31ab340a3251493a0683262454452f --- /dev/null +++ b/elegance-niji/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a35777fa3f585adf73d3f73fce45335386528ccd5b01a49430faca4f4e5bb335 +size 859528 diff --git a/elegance-niji/example_images/04_image.jpeg b/elegance-niji/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..a667707e3975fbf21311659f83595c489ea56870 --- /dev/null +++ b/elegance-niji/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:addcbd1ca0d11c541c8e37aa4cd236d420c80d30fe4fb537c72654f603700860 +size 1322149 diff --git a/elegance-niji/example_images/05_image.jpeg b/elegance-niji/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..f3ba47cc0a32799f06c3b7282ecdc3077ccf8890 --- /dev/null +++ b/elegance-niji/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1c47f4e0b7ca8e7111eba526eaaf15c7774155c27ba872f189aba3097f6f17bb +size 2243047 diff --git a/elegance-niji/example_images/06_image.jpeg b/elegance-niji/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..11eb56ad882e7102314b1cf55cf43a8ef6cb585c --- /dev/null +++ b/elegance-niji/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:15081d060b4d10d7ea569b86687d23d5c1a1458bb675ae44610aecc475f0f130 +size 438270 diff --git a/elegance-niji/info.txt b/elegance-niji/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..11db2bbe18d39e3e8cc4bf923f03f20c0365068a --- /dev/null +++ b/elegance-niji/info.txt @@ -0,0 +1,264 @@ +═══════════════════════════════════════════════════════════════════════════════ + ELEGANCE NIJI +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Visual style + Author purplelady + Source https://civitai.com/models/1520081 + Provenance CivitAI · model 1520081 · version 3209626 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File Elegance_Niji_KR2.safetensors + Version v1.0 Krea 2 + Size 218.0 MB + Format SafeTensor / unknown precision + SHA256 1a71faadb8a68220eb28883ee7ef7d0317a63b0995ce45a76fe927cd99675ceb + Published 2026-08-08 + +─────────────────────────────────────────────────────────────────────────────── + 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 +─────────────────────────────────────────────────────────────────────────────── + Taken verbatim from the author's description: + + LoRA strength 0.8–1.2 + +─────────────────────────────────────────────────────────────────────────────── + ADOPTION & POPULARITY +─────────────────────────────────────────────────────────────────────────────── + Downloads 4,628 + Rating 100% positive (628/628) + Comments 11 + Published 27 days ago + Download rate 50.7/day since release + + Adoption rank #122 of 150 in this collection (Niche) + Momentum rank #50 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 — 4,628 downloads. [source: CivitAI stats] + + Positively received — 628 thumbs-up with no down-votes. [source: + CivitAI stats] + + Compact at 218 MB — cheap to stack with other LoRAs. [source: file + size] + + Author published concrete recommended settings. [source: description] + +─────────────────────────────────────────────────────────────────────────────── + CONS & CAVEATS +─────────────────────────────────────────────────────────────────────────────── + - 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 +─────────────────────────────────────────────────────────────────────────────── + Tigger word: 3l3ganc3 ZIT Weights : 0.8 -1.2 + So added the ZIT version, I am very happy with the results even with + training on a much smaller data set, there as still a few issues and my + prompting style is a mess (sorry i am new to natrual language prompting) + one of the things i am trying to fix in possible future versions is + reducing the very warm coloring this lora defaults to (you can brute force + it with prompts if you want). Please feel free to share your feedback, I am + still trying to refine it :) ---------------------------------------------- + -------------------------------------------------- Illustrious Weights : 1 + -1.6 This is lora was designed to replicate some the niji 6 realistic + anime/draw art style, the data set mostly focused on portraits for male and + female but you can brute force it for more dramatic poses. + + Some of the features can be triggers using prompts like " 3l3ganc3, + realistic, semi-realistic " etc. . + + I have tried it with several illustrious/SDXL/Pony models and it seem to + works well flat anime style as well as more 3d styles (though it intended + more for the former, I shall update more resources that work with it as I + discover them myself). + +─────────────────────────────────────────────────────────────────────────────── + 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, 1152x1536) + 02_image.jpeg (image, 2560x3840) + 03_image.jpeg (image, 1662x2216) + 04_image.jpeg (image, 888x1184) + 05_image.jpeg (image, 1024x1312) + 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 00:16 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/elegance-niji/metadata.json b/elegance-niji/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..85b98e4ce8dc0dae10d9bf3eaf3c60ef099be078 --- /dev/null +++ b/elegance-niji/metadata.json @@ -0,0 +1,133 @@ +{ + "name": "Elegance Niji", + "slug": "elegance-niji", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 1520081, + "version_id": 3209626, + "version_name": "v1.0 Krea 2", + "source": "civitai", + "url": "https://civitai.com/models/1520081", + "repo_id": null, + "file_path": null, + "author": "purplelady", + "category": { + "key": "style", + "label": "Visual style" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [], + "recommended_settings": { + "LoRA strength": "0.8–1.2" + }, + "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 — 4,628 downloads. [source: CivitAI stats]", + "Positively received — 628 thumbs-up with no down-votes. [source: CivitAI stats]", + "Compact at 218 MB — cheap to stack with other LoRAs. [source: file size]", + "Author published concrete recommended settings. [source: description]" + ], + "cons": [ + "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": "Elegance_Niji_KR2.safetensors", + "size_bytes": 228588856, + "size_mb": 218.0, + "sha256": "1a71faadb8a68220eb28883ee7ef7d0317a63b0995ce45a76fe927cd99675ceb" + }, + "stats": { + "downloads": 4628, + "thumbs_up": 628, + "thumbs_down": 0 + }, + "popularity": { + "downloads": 4628, + "thumbs_up": 628, + "thumbs_down": 0, + "comments": 11, + "approval": "100% positive (628/628)", + "age_days": 27.5, + "downloads_per_day": 50.75, + "rank_by_downloads": 122, + "rank_by_momentum": 50, + "of_total": 150, + "adoption_tier": "Niche", + "momentum": "Steady · fresh release" + }, + "selection_score": 12164.0, + "selection_reasons": [ + "4,628 downloads", + "628 thumbs-up", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "1152x1536", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "2560x3840", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "1662x2216", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "888x1184", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "1024x1312", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "832x1248", + "prompt": "" + } + ], + "tags": [ + "style", + "semi-realistic" + ], + "scraped_at": "2026-09-05T00:16:08+00:00", + "path": "krea-2/elegance-niji" +} diff --git a/glowing-and-light-particles/example_images/01_image.jpeg b/glowing-and-light-particles/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..26b44c3c4eed0ef683d6f3561ce5673b610ef100 --- /dev/null +++ b/glowing-and-light-particles/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80d16ddac7d3ad5211a2ad6c7a3e5704b2c351f2030cd42b3d36ad3cee70d068 +size 7583295 diff --git a/glowing-and-light-particles/example_images/02_image.jpeg b/glowing-and-light-particles/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..59be0bf509cebc220c8ff26c21edcf0e5560bf6c --- /dev/null +++ b/glowing-and-light-particles/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9886630c8001ef8037f10d2ebb4c2fdfef0a679c89bd3c7ea5cde417b52b0d4a +size 610118 diff --git a/glowing-and-light-particles/example_images/03_image.jpeg b/glowing-and-light-particles/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..ff68479d7a0ef53468f10fbc9f5db78fd74876c6 --- /dev/null +++ b/glowing-and-light-particles/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:44adbf009dc775297321db7c2f5319cd8e18b010013aaf9c1bee5625744b44f2 +size 5306225 diff --git a/glowing-and-light-particles/example_images/04_image.jpeg b/glowing-and-light-particles/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..a5a8c6f30b4bf50f20d4a8926890119fe8f3f7e3 --- /dev/null +++ b/glowing-and-light-particles/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:858f10f08e9d5f94138ee4d243e631bab27e7e84dfa54d1b3d70de3544595123 +size 7975972 diff --git a/glowing-and-light-particles/example_images/05_image.jpeg b/glowing-and-light-particles/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..c254706f26432229d0d8b601f8bc0fa83b39b09d --- /dev/null +++ b/glowing-and-light-particles/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d12f1a04c7c7895f0594a145abe97114da047fb22f0bea394f30644da5c091a7 +size 361674 diff --git a/glowing-and-light-particles/example_images/06_image.jpeg b/glowing-and-light-particles/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..f1b9d5cd64f0e9c479b295f1cb867f4f145804d6 --- /dev/null +++ b/glowing-and-light-particles/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ff0ef6f3c5d3986232a8fc7983b0160d856b3f7a9817a2f98c6a3f67588564af +size 7114499 diff --git a/glowing-and-light-particles/glowing_epoch_10.safetensors b/glowing-and-light-particles/glowing_epoch_10.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..5289c373b8ac61c3e3bebf8f4878291281c736ae --- /dev/null +++ b/glowing-and-light-particles/glowing_epoch_10.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:12d700dc128da0c1883fe8b14c70a080d9ee0e5bb6cf43432d1a362e292da651 +size 228588912 diff --git a/glowing-and-light-particles/info.txt b/glowing-and-light-particles/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..2d04ae888f89aa67f674287e9b7a5db93d91ff06 --- /dev/null +++ b/glowing-and-light-particles/info.txt @@ -0,0 +1,276 @@ +═══════════════════════════════════════════════════════════════════════════════ + GLOWING AND LIGHT PARTICLES +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Uncategorised + Author zwaetschge + Source https://civitai.com/models/573849 + Provenance CivitAI · model 573849 · version 3216784 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File glowing_epoch_10.safetensors + Version v1.0 K2 + Size 218.0 MB + Format SafeTensor / unknown precision + SHA256 12d700dc128da0c1883fe8b14c70a080d9ee0e5bb6cf43432d1a362e292da651 + Published 2026-08-10 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + › "glowing" + +─────────────────────────────────────────────────────────────────────────────── + 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 5,708 + Rating 100% positive (555/555) + Comments 0 + Published 25 days ago + Download rate 16.0/day since release + + Adoption rank #103 of 150 in this collection (Modest) + Momentum rank #113 of 150 by download rate (Long tail · fresh release) + + Adoption tier is the percentile of total downloads within this collection; + momentum is the percentile of downloads-per-day since release. Momentum is + ranked rather than measured against a fixed rate, because this base model + is itself new and every LoRA looks fast on an absolute scale. Both numbers + are CivitAI's, read on the scrape date at the foot of this file. + +─────────────────────────────────────────────────────────────────────────────── + CHAINING / STACKING +─────────────────────────────────────────────────────────────────────────────── + Stacks with others not stated + + The author gave no chaining guidance. The rules below are the + general Krea 2 ones, not anything specific to this LoRA. + + General Krea 2 rules that apply here (see REFERENCE.txt): + + • The most important rule is not about chaining but about which + checkpoint you chain onto: the official README says in bold "TRAIN on + Raw and RUN on Turbo", and "We highly recommend using RAW for training + LoRAs and applying them on Turbo for inference." LoRAs trained on Raw + are stated to work well on Turbo. + • Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 + LoRAs are DiT-only, so there is no CLIP side to chain and the text- + encoder branch stays untouched however many LoRAs you add. + • Krea 2 LoRAs stack in practice and the community treats multi-LoRA as + routine - style + character, style + slider, realism + character. + Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one + writes "I recently stacked over 20 of my own LoRAs on a single + generation and the output still came out great." + • Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge + on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is + loaded. The CivitAI 12GB training guide also notes rank affects + composability: "rank 8 produced a lighter style influence that was + easier to stack with other LoRAs, rank 16 provided a useful balance + between concept strength and composability, and rank 32 was better + suited to a precise object or concept intended to take priority over + other stacked LoRAs." + • Accelerator LoRAs are position-sensitive in intent, not in node order: + a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a + sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to- + Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is + explicitly recommended by the 4-step LoRA's author. + • Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI + (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card + says "Keep mu = 1.15. The training targets are anchored to that grid; a + different shift moves the [sampler off them]"; the same reasoning + applies to any Turbo-trained adapter in the chain. + • Reference/edit LoRAs are designed to be stacked under ordinary subject + and style LoRAs - the Identity Edit card states "Composes with your + LoRAs: character/body/style LoRAs stack on top and steer the prior" and + recommends a stacked subject LoRA when unusual subjects drift toward + the base prior. + • Krea 2 LoRAs cannot be chained with LoRAs for any other base model. + Many CivitAI pages carry multi-model titles like "(Krea-2 + ZIT)" or + "[Flux | ZIB | Krea2]" - those are separate files for separate + architectures shipped under one model page, not a combined adapter. + • To combine two full Krea 2 checkpoints rather than adapters, ComfyUI + ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per- + section weights for first., tmlp., txtmlp., tproj., + txtfusion.projector, txtfusion.layerwise_blocks.0-1, + txtfusion.refiner_blocks.0-1, blocks.0-27 and last. + • The author never states which weight family this targets. Silent + degradation, never a hard error, between Raw and Turbo. The two share + one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" + - so every LoRA tensor shape matches on both and ComfyUI/diffuser + +─────────────────────────────────────────────────────────────────────────────── + PROS +─────────────────────────────────────────────────────────────────────────────── + + Widely adopted — 5,708 downloads. [source: CivitAI stats] + + Positively received — 555 thumbs-up with no down-votes. [source: + CivitAI stats] + + Explicit trigger word(s) — glowing — 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] + - 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 +─────────────────────────────────────────────────────────────────────────────── + This model enables objects to emit a glowing effect or appear as if + composed of light particles. + +─────────────────────────────────────────────────────────────────────────────── + 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, 2688x2160) + 02_image.jpeg (image, 1023x1537) + 03_image.jpeg (image, 2688x2160) + 04_image.jpeg (image, 1492x2466) + 05_image.jpeg (image, 832x1248) + 06_image.jpeg (image, 2688x2160) + +─────────────────────────────────────────────────────────────────────────────── + 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 00:15 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/glowing-and-light-particles/metadata.json b/glowing-and-light-particles/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..548feb6fea94c85bab1c1c96849ce4a4fe960435 --- /dev/null +++ b/glowing-and-light-particles/metadata.json @@ -0,0 +1,132 @@ +{ + "name": "Glowing and Light Particles", + "slug": "glowing-and-light-particles", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 573849, + "version_id": 3216784, + "version_name": "v1.0 K2", + "source": "civitai", + "url": "https://civitai.com/models/573849", + "repo_id": null, + "file_path": null, + "author": "zwaetschge", + "category": { + "key": "other", + "label": "Uncategorised" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [ + "glowing" + ], + "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 — 5,708 downloads. [source: CivitAI stats]", + "Positively received — 555 thumbs-up with no down-votes. [source: CivitAI stats]", + "Explicit trigger word(s) — glowing — 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]", + "Thin documentation — expect trial and error. [source: description length]" + ], + "file": { + "name": "glowing_epoch_10.safetensors", + "size_bytes": 228588912, + "size_mb": 218.0, + "sha256": "12d700dc128da0c1883fe8b14c70a080d9ee0e5bb6cf43432d1a362e292da651" + }, + "stats": { + "downloads": 5708, + "thumbs_up": 555, + "thumbs_down": 0 + }, + "popularity": { + "downloads": 5708, + "thumbs_up": 555, + "thumbs_down": 0, + "comments": 0, + "approval": "100% positive (555/555)", + "age_days": 25.2, + "downloads_per_day": 16.03, + "rank_by_downloads": 103, + "rank_by_momentum": 113, + "of_total": 150, + "adoption_tier": "Modest", + "momentum": "Long tail · fresh release" + }, + "selection_score": 12368.0, + "selection_reasons": [ + "5,708 downloads", + "555 thumbs-up", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "2688x2160", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "1023x1537", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "2688x2160", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "1492x2466", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "832x1248", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "2688x2160", + "prompt": "" + } + ], + "tags": [ + "concept" + ], + "scraped_at": "2026-09-05T00:15:42+00:00", + "path": "krea-2/glowing-and-light-particles" +} diff --git a/harness-panel-gag/Harness_Panel_Gag_Krea_2_epoch_7.safetensors b/harness-panel-gag/Harness_Panel_Gag_Krea_2_epoch_7.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..622f586520098fbd182b215b109f2e67f06a0f34 --- /dev/null +++ b/harness-panel-gag/Harness_Panel_Gag_Krea_2_epoch_7.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1af244f61ee16c78176f6eddf340bc840dd8a764f9b035cea9b0330daa9be12b +size 228588904 diff --git a/harness-panel-gag/example_images/01_image.jpeg b/harness-panel-gag/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..f3cbe3f663a04ffd22adea5898ad9ed985a6a8cd --- /dev/null +++ b/harness-panel-gag/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b279fd125ef1867f6c400717b91b5ce9e4032c983fc899afb0163e2b19565882 +size 1272012 diff --git a/harness-panel-gag/example_images/02_image.jpeg b/harness-panel-gag/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..db7f10126dba89c6e2d7902d7503176ead19d999 --- /dev/null +++ b/harness-panel-gag/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:36739dcf411a08e84843f0a2a5be3e202ad1a452f8aec5082a373111c080d082 +size 1085462 diff --git a/harness-panel-gag/example_images/03_image.jpeg b/harness-panel-gag/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..99dd4ad19eed1b6e4dd89f4ddea45a28a2c599fe --- /dev/null +++ b/harness-panel-gag/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cdce0b67d1cbc0405b576da821e81d09bc930248e9ce253752a6d7840d6bf92c +size 1335067 diff --git a/harness-panel-gag/example_images/04_image.jpeg b/harness-panel-gag/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..8846315965e294e5165c06529048e533fcf42f91 --- /dev/null +++ b/harness-panel-gag/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4671eb0558e4d37d488505828d1154a05580c7f7afa063db0f722ffc34cfc4b8 +size 1303791 diff --git a/harness-panel-gag/example_images/05_image.jpeg b/harness-panel-gag/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..065c9e354b5535358f5207fd10663aff560f2c29 --- /dev/null +++ b/harness-panel-gag/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dabfa4f1053adca724d6653d0887ae8652f3fc80ee654ca8de795215d35480d2 +size 1025727 diff --git a/harness-panel-gag/example_images/06_image.jpeg b/harness-panel-gag/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..9900117428d1a40fccb082d44ef152d1795165f2 --- /dev/null +++ b/harness-panel-gag/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:88bcde214addd56593991399f15c660b8dc236f52a2c6facd7d03b0862eb85d8 +size 1060363 diff --git a/harness-panel-gag/info.txt b/harness-panel-gag/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..6bb0b0e290f3be5681f09afe09967d68c5564979 --- /dev/null +++ b/harness-panel-gag/info.txt @@ -0,0 +1,272 @@ +═══════════════════════════════════════════════════════════════════════════════ + HARNESS PANEL GAG +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Visual style + Author John_KSampler + Source https://civitai.com/models/1209663 + Provenance CivitAI · model 1209663 · version 3142354 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File Harness_Panel_Gag_Krea_2_epoch_7.safetensors + Version KREA 2 + Size 218.0 MB + Format SafeTensor / unknown precision + SHA256 1af244f61ee16c78176f6eddf340bc840dd8a764f9b035cea9b0330daa9be12b + Published 2026-07-18 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + › "harness panel gag over mouth" + +─────────────────────────────────────────────────────────────────────────────── + RECOMMENDED SETTINGS +─────────────────────────────────────────────────────────────────────────────── + Taken verbatim from the author's description: + + LoRA strength 1.0 + +─────────────────────────────────────────────────────────────────────────────── + ADOPTION & POPULARITY +─────────────────────────────────────────────────────────────────────────────── + Downloads 5,912 + Rating 100% positive (423/423) + Comments 0 + Published 48 days ago + Download rate 10.9/day since release + + Adoption rank #96 of 150 in this collection (Modest) + Momentum rank #137 of 150 by download rate (Long tail) + + Adoption tier is the percentile of total downloads within this collection; + momentum is the percentile of downloads-per-day since release. Momentum is + ranked rather than measured against a fixed rate, because this base model + is itself new and every LoRA looks fast on an absolute scale. Both numbers + are CivitAI's, read on the scrape date at the foot of this file. + +─────────────────────────────────────────────────────────────────────────────── + CHAINING / STACKING +─────────────────────────────────────────────────────────────────────────────── + Stacks with others not stated + + The author gave no chaining guidance. The rules below are the + general Krea 2 ones, not anything specific to this LoRA. + + General Krea 2 rules that apply here (see REFERENCE.txt): + + • The most important rule is not about chaining but about which + checkpoint you chain onto: the official README says in bold "TRAIN on + Raw and RUN on Turbo", and "We highly recommend using RAW for training + LoRAs and applying them on Turbo for inference." LoRAs trained on Raw + are stated to work well on Turbo. + • Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 + LoRAs are DiT-only, so there is no CLIP side to chain and the text- + encoder branch stays untouched however many LoRAs you add. + • Krea 2 LoRAs stack in practice and the community treats multi-LoRA as + routine - style + character, style + slider, realism + character. + Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one + writes "I recently stacked over 20 of my own LoRAs on a single + generation and the output still came out great." + • Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge + on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is + loaded. The CivitAI 12GB training guide also notes rank affects + composability: "rank 8 produced a lighter style influence that was + easier to stack with other LoRAs, rank 16 provided a useful balance + between concept strength and composability, and rank 32 was better + suited to a precise object or concept intended to take priority over + other stacked LoRAs." + • Accelerator LoRAs are position-sensitive in intent, not in node order: + a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a + sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to- + Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is + explicitly recommended by the 4-step LoRA's author. + • Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI + (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card + says "Keep mu = 1.15. The training targets are anchored to that grid; a + different shift moves the [sampler off them]"; the same reasoning + applies to any Turbo-trained adapter in the chain. + • Reference/edit LoRAs are designed to be stacked under ordinary subject + and style LoRAs - the Identity Edit card states "Composes with your + LoRAs: character/body/style LoRAs stack on top and steer the prior" and + recommends a stacked subject LoRA when unusual subjects drift toward + the base prior. + • Krea 2 LoRAs cannot be chained with LoRAs for any other base model. + Many CivitAI pages carry multi-model titles like "(Krea-2 + ZIT)" or + "[Flux | ZIB | Krea2]" - those are separate files for separate + architectures shipped under one model page, not a combined adapter. + • To combine two full Krea 2 checkpoints rather than adapters, ComfyUI + ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per- + section weights for first., tmlp., txtmlp., tproj., + txtfusion.projector, txtfusion.layerwise_blocks.0-1, + txtfusion.refiner_blocks.0-1, blocks.0-27 and last. + • The author never states which weight family this targets. Silent + degradation, never a hard error, between Raw and Turbo. The two share + one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" + - so every LoRA tensor shape matches on both and ComfyUI/diffuser + +─────────────────────────────────────────────────────────────────────────────── + PROS +─────────────────────────────────────────────────────────────────────────────── + + Widely adopted — 5,912 downloads. [source: CivitAI stats] + + Positively received — 423 thumbs-up with no down-votes. [source: + CivitAI stats] + + Explicit trigger word(s) — harness panel gag over mouth — 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 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 +─────────────────────────────────────────────────────────────────────────────── + I had been waiting for this from the usual suspects, since SDXL and Pony + version do exist, but it appears to not be happening any time soon. I guess + if you want something done... + So, basic harness panel gag for Illustrious base models. Technically SFW + but... you know. If you don't, ask your mom. Illustrious can already do the + concept(s) ("panel gag" and " face gag harness") but it's not reliable, + plus it really likes to put holes everywhere, which means that even if you + use an SDXL LoRA, you will more likely than not end up with something like + a plug gag . This LoRA is intended to fix this issue. It appears that it + still wants to grab on to anything it can in the center, especially in + close-ups. Trained on a mix of photos, illustrations and hand-picked AI + images with relevant photoshop applied where applicable. Unlike some of the + existing purely photo-based ones, it won't pull your image massively + towards realism. That last point is also the reason there's a Pony version + of it now (apart from it being easy since I already had the dataset, + naturally): It's more docile than the existing LoRAs if you want + source_anime or source_cartoon, but a lot less so if you want realism. The + SDXL version is pretty much just to complete the set, but it works well + with Animagine and other illustration-focused checkpoints. + Recommended prompt is " panelgag, panel gag " and " face harness " for + v0.5/" gag harness " for v1.0. " covered mouth " can help if you still get + mouth artifacts. Strength 1.0 seems to work fine, go lower for close-ups + and higher for wide shots if necessary as per usual. The " gag harness " + tag has a tendency to imply body harnesses, put " breast harness, bondage + harness " into the negatives if that's a problem. You can also try leaving + the " gag harness " prompt out completely, which is especially helpful if + you want it to be partially covered by hair or headwear of some kind. + You can choose any color, as long as it's black. Turns out color can be + done, it just might require upweighting of the color prompt. + +─────────────────────────────────────────────────────────────────────────────── + 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, 832x1216) + 02_image.jpeg (image, 832x1216) + 03_image.jpeg (image, 832x1216) + 04_image.jpeg (image, 832x1216) + 05_image.jpeg (image, 832x1216) + 06_image.jpeg (image, 832x1216) + +─────────────────────────────────────────────────────────────────────────────── + 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 00:16 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/harness-panel-gag/metadata.json b/harness-panel-gag/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..2b206d519f84921752691698af3bc0756b48e80e --- /dev/null +++ b/harness-panel-gag/metadata.json @@ -0,0 +1,142 @@ +{ + "name": "Harness Panel Gag", + "slug": "harness-panel-gag", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 1209663, + "version_id": 3142354, + "version_name": "KREA 2", + "source": "civitai", + "url": "https://civitai.com/models/1209663", + "repo_id": null, + "file_path": null, + "author": "John_KSampler", + "category": { + "key": "style", + "label": "Visual style" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [ + "harness panel gag over mouth" + ], + "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 — 5,912 downloads. [source: CivitAI stats]", + "Positively received — 423 thumbs-up with no down-votes. [source: CivitAI stats]", + "Explicit trigger word(s) — harness panel gag over mouth — 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 never states which generation mode it was trained for; test against your own workflow before relying on it. [source: description]" + ], + "file": { + "name": "Harness_Panel_Gag_Krea_2_epoch_7.safetensors", + "size_bytes": 228588904, + "size_mb": 218.0, + "sha256": "1af244f61ee16c78176f6eddf340bc840dd8a764f9b035cea9b0330daa9be12b" + }, + "stats": { + "downloads": 5912, + "thumbs_up": 423, + "thumbs_down": 0 + }, + "popularity": { + "downloads": 5912, + "thumbs_up": 423, + "thumbs_down": 0, + "comments": 0, + "approval": "100% positive (423/423)", + "age_days": 48.0, + "downloads_per_day": 10.89, + "rank_by_downloads": 96, + "rank_by_momentum": 137, + "of_total": 150, + "adoption_tier": "Modest", + "momentum": "Long tail" + }, + "selection_score": 11388.0, + "selection_reasons": [ + "5,912 downloads", + "423 thumbs-up", + "documented triggers + usage notes", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "832x1216", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "832x1216", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "832x1216", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "832x1216", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "832x1216", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "832x1216", + "prompt": "" + } + ], + "tags": [ + "leather", + "gear", + "harness", + "gag", + "illustration", + "bondage", + "clothing", + "bdsm" + ], + "scraped_at": "2026-09-05T00:16:21+00:00", + "path": "krea-2/harness-panel-gag" +} diff --git a/krea-2-realism-slider/RealismSlider-v1.safetensors b/krea-2-realism-slider/RealismSlider-v1.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..13bed995ee65ccf4e1a1d338273d5a8f55cab01c --- /dev/null +++ b/krea-2-realism-slider/RealismSlider-v1.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:89744898461de18d91f26c6b368872526df462a74a9243e70253f40a677950b8 +size 6719408 diff --git a/krea-2-realism-slider/example_images/01_image.jpeg b/krea-2-realism-slider/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..7b8d0d97d16570b31533f59093d8f4af980a31a4 --- /dev/null +++ b/krea-2-realism-slider/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:47e3db074449edc07c915e2a9a70a5ddc063350aa91415e2f89cbc86312db1f0 +size 2636711 diff --git a/krea-2-realism-slider/example_images/02_image.jpeg b/krea-2-realism-slider/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..5f5d0b1f597d7e18e2f9a9a703e43b83e469477a --- /dev/null +++ b/krea-2-realism-slider/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4fed36e7df5e3ce21209a2430df4a0f94c31a81e856d165fd4f2951389977c8b +size 123574 diff --git a/krea-2-realism-slider/example_images/03_image.jpeg b/krea-2-realism-slider/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..bb133fcd6f3553cb5f2e11102e4236276a7ff9b9 --- /dev/null +++ b/krea-2-realism-slider/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:271749b7ebb4fca465013c1daf3bec4842f80eff732181e3eeab3e3e300c8497 +size 315507 diff --git a/krea-2-realism-slider/example_images/04_image.jpeg b/krea-2-realism-slider/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..4102e234c415b873308cdc80b3457c5bea678588 --- /dev/null +++ b/krea-2-realism-slider/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d8193883109d11e89b36e7778f449e64719a56ee99919da66038ee37db740cfb +size 279752 diff --git a/krea-2-realism-slider/example_images/05_image.jpeg b/krea-2-realism-slider/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..72dc717eab2210628eb8f6ffa1b2040405a25e32 --- /dev/null +++ b/krea-2-realism-slider/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fc5ba0dc539959eb366542cd218071f2ec080943f8f8cab8e7567e9d623c37c1 +size 13226381 diff --git a/krea-2-realism-slider/example_images/06_image.jpeg b/krea-2-realism-slider/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..0df7e9ef730fe64ae88286932704bf8aa5f2cd67 --- /dev/null +++ b/krea-2-realism-slider/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a193a2f17fb3af837391b352586d13ab1ada1a65c76d72d3422867268fdbe669 +size 6373517 diff --git a/krea-2-realism-slider/info.txt b/krea-2-realism-slider/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..76838798e5976012f17c4aae9502ac8fd8bd25f9 --- /dev/null +++ b/krea-2-realism-slider/info.txt @@ -0,0 +1,293 @@ +═══════════════════════════════════════════════════════════════════════════════ + [KREA 2] REALISM SLIDER +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Photorealism + Author alcaitiff + Source https://civitai.com/models/2781697 + Provenance CivitAI · model 2781697 · version 3132956 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File RealismSlider-v1.safetensors + Version v1.0 + Size 6.4 MB + Format SafeTensor / unknown precision + SHA256 89744898461de18d91f26c6b368872526df462a74a9243e70253f40a677950b8 + Published 2026-07-15 + +─────────────────────────────────────────────────────────────────────────────── + 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 7,042 + Rating 100% positive (396/397) + Comments 12 + Published 51 days ago + Download rate 137.5/day since release + + Adoption rank #83 of 150 in this collection (Modest) + Momentum rank #16 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 — 7,042 downloads. [source: CivitAI stats] + + Positively received — 396 thumbs-up, 1 down. [source: CivitAI stats] + + Compact at 6 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 +─────────────────────────────────────────────────────────────────────────────── + Negative values push the image toward illustration, positive values push it + toward realism. + Krea 2 is already ridiculously good at styles it knows, so most of the time + the "correct" solution is probably a better prompt. That said, sometimes + you just want to type a character name, throw a LoRA on it, and move on + with your life. That's where this slider comes in. + Think of it as a style steering wheel. If the model is looking a little too + much like concept art, push it toward realism. If it's getting too + photographic and losing the illustrated charm, pull it back the other way. + Values around -1 to +1 are usually the sweet spot. You can absolutely go + further if you want stronger results, but at some point you're basically + telling the model to stop negotiating and commit to one side. + I've included a few examples because the effect is easier to see than + explain. The slider doesn't just add detail or remove it—it changes how + Krea interprets the image, from anime eyes and painted textures all the way + to skin pores and camera-like realism. Use it whenever the model's idea and + your idea aren't quite having the same conversation. + +─────────────────────────────────────────────────────────────────────────────── + VERSION NOTES +─────────────────────────────────────────────────────────────────────────────── + First 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, 1024x1536) + 02_image.jpeg (image, 832x1248) + 03_image.jpeg (image, 768x1376) + 04_image.jpeg (image, 832x1248) + 05_image.jpeg (image, 2496x3344) + 06_image.jpeg (image, 2048x2048) + +─────────────────────────────────────────────────────────────────────────────── + 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 00:16 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/krea-2-realism-slider/metadata.json b/krea-2-realism-slider/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..c7fbea55eebf400ef214a786e2aa6ae635587e25 --- /dev/null +++ b/krea-2-realism-slider/metadata.json @@ -0,0 +1,131 @@ +{ + "name": "[KREA 2] Realism Slider", + "slug": "krea-2-realism-slider", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 2781697, + "version_id": 3132956, + "version_name": "v1.0", + "source": "civitai", + "url": "https://civitai.com/models/2781697", + "repo_id": null, + "file_path": null, + "author": "alcaitiff", + "category": { + "key": "photoreal", + "label": "Photorealism" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [], + "recommended_settings": {}, + "chaining": { + "author": { + "stacks": "not stated", + "notes": "", + "avoid": [], + "requires": [], + "order": "", + "strength": "", + "stated": false + }, + "general_rules": [ + "The most important rule is not about chaining but about which checkpoint you chain onto: the official README says in bold \"TRAIN on Raw and RUN on Turbo\", and \"We highly recommend using RAW for training LoRAs and applying them on Turbo for inference.\" LoRAs trained on Raw are stated to work well on Turbo.", + "Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 LoRAs are DiT-only, so there is no CLIP side to chain and the text-encoder branch stays untouched however many LoRAs you add.", + "Krea 2 LoRAs stack in practice and the community treats multi-LoRA as routine - style + character, style + slider, realism + character. Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one writes \"I recently stacked over 20 of my own LoRAs on a single generation and the output still came out great.\"", + "Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is loaded. The CivitAI 12GB training guide also notes rank affects composability: \"rank 8 produced a lighter style influence that was easier to stack with other LoRAs, rank 16 provided a useful balance between concept strength and composability, and rank 32 was better suited to a precise object or concept intended to take priority over other stacked LoRAs.\"", + "Accelerator LoRAs are position-sensitive in intent, not in node order: a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to-Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is explicitly recommended by the 4-step LoRA's author.", + "Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card says \"Keep mu = 1.15. The training targets are anchored to that grid; a different shift moves the [sampler off them]\"; the same reasoning applies to any Turbo-trained adapter in the chain.", + "Reference/edit LoRAs are designed to be stacked under ordinary subject and style LoRAs - the Identity Edit card states \"Composes with your LoRAs: character/body/style LoRAs stack on top and steer the prior\" and recommends a stacked subject LoRA when unusual subjects drift toward the base prior.", + "Krea 2 LoRAs cannot be chained with LoRAs for any other base model. Many CivitAI pages carry multi-model titles like \"(Krea-2 + ZIT)\" or \"[Flux | ZIB | Krea2]\" - those are separate files for separate architectures shipped under one model page, not a combined adapter.", + "To combine two full Krea 2 checkpoints rather than adapters, ComfyUI ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per-section weights for first., tmlp., txtmlp., tproj., txtfusion.projector, txtfusion.layerwise_blocks.0-1, txtfusion.refiner_blocks.0-1, blocks.0-27 and last.", + "The author never states which weight family this targets. Silent degradation, never a hard error, between Raw and Turbo. The two share one architecture - ai-toolkit's config is commented \"The reference 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)\" - so every LoRA tensor shape matches on both and ComfyUI/diffuser" + ] + }, + "pros": [ + "Widely adopted — 7,042 downloads. [source: CivitAI stats]", + "Positively received — 396 thumbs-up, 1 down. [source: CivitAI stats]", + "Compact at 6 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": "RealismSlider-v1.safetensors", + "size_bytes": 6719408, + "size_mb": 6.41, + "sha256": "89744898461de18d91f26c6b368872526df462a74a9243e70253f40a677950b8" + }, + "stats": { + "downloads": 7042, + "thumbs_up": 396, + "thumbs_down": 1 + }, + "popularity": { + "downloads": 7042, + "thumbs_up": 396, + "thumbs_down": 1, + "comments": 12, + "approval": "100% positive (396/397)", + "age_days": 51.2, + "downloads_per_day": 137.55, + "rank_by_downloads": 83, + "rank_by_momentum": 16, + "of_total": 150, + "adoption_tier": "Modest", + "momentum": "Trending" + }, + "selection_score": 11794.0, + "selection_reasons": [ + "7,042 downloads", + "396 thumbs-up", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "1024x1536", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "832x1248", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "768x1376", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "832x1248", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "2496x3344", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "2048x2048", + "prompt": "" + } + ], + "tags": [ + "realism", + "tool", + "illustration" + ], + "scraped_at": "2026-09-05T00:16:04+00:00", + "path": "krea-2/krea-2-realism-slider" +} diff --git a/krea2-qwen2511-z-image-illustria-anime-01/example_images/01_image.jpeg b/krea2-qwen2511-z-image-illustria-anime-01/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..b607efb42e0e98f6db014c7c0d0df855a6666e77 --- /dev/null +++ b/krea2-qwen2511-z-image-illustria-anime-01/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6cceddda47f9e90c8eda747adf04e077d45182d3af90ad05e2b913c31a5ed546 +size 8315457 diff --git a/krea2-qwen2511-z-image-illustria-anime-01/example_images/02_image.jpeg b/krea2-qwen2511-z-image-illustria-anime-01/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..b77f76eab5210278ec23ecef07223a75abe1c2e2 --- /dev/null +++ b/krea2-qwen2511-z-image-illustria-anime-01/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e42ea7297d76013878946daf8b9187c9fee3600058ddc1909e05d993a4c10e61 +size 9020237 diff --git a/krea2-qwen2511-z-image-illustria-anime-01/example_images/03_image.jpeg b/krea2-qwen2511-z-image-illustria-anime-01/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..8a5c58b4698a7bd5c915348f582c51f700079082 --- /dev/null +++ b/krea2-qwen2511-z-image-illustria-anime-01/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:847c3258f1e3e677b9b1dc5c6ff4cac7c962bdeb7096f2e4a5957c0e000ad44a +size 1310980 diff --git a/krea2-qwen2511-z-image-illustria-anime-01/example_images/04_image.jpeg b/krea2-qwen2511-z-image-illustria-anime-01/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..bfd0a38922e783ddd94175be96a37417b1e5cd8a --- /dev/null +++ b/krea2-qwen2511-z-image-illustria-anime-01/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d7a03e242dd07b7a84f795b9140c4a20067a0eb1e3a595ab5b4d052b748b04a +size 655509 diff --git a/krea2-qwen2511-z-image-illustria-anime-01/example_images/05_image.jpeg b/krea2-qwen2511-z-image-illustria-anime-01/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..a562f742d02241de089cda47ae63b0fb8204f5c5 --- /dev/null +++ b/krea2-qwen2511-z-image-illustria-anime-01/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b00eaf704a0a5acd559dd6b0431d2e00881136efb05352a58a7e68ecd32d3b89 +size 1285466 diff --git a/krea2-qwen2511-z-image-illustria-anime-01/example_images/06_image.jpeg b/krea2-qwen2511-z-image-illustria-anime-01/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..19e661d02796585d267c3d0d2b6bcc516c5e73dc --- /dev/null +++ b/krea2-qwen2511-z-image-illustria-anime-01/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:49508114df41a416a8f867ac2b73c3c3dcd33f1527db9833ff70ae544e159d65 +size 251060 diff --git a/krea2-qwen2511-z-image-illustria-anime-01/info.txt b/krea2-qwen2511-z-image-illustria-anime-01/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..ea91dc390ac208d5938cff23114675d90de4da1a --- /dev/null +++ b/krea2-qwen2511-z-image-illustria-anime-01/info.txt @@ -0,0 +1,275 @@ +═══════════════════════════════════════════════════════════════════════════════ + KREA2+QWEN2511+Z-IMAGE ILLUSTRIA ANIME 01 +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Visual style + Author _Envy_ + Source https://civitai.com/models/2174309 + Provenance CivitAI · model 2174309 · version 3072531 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File k2-illustria3_000000400.safetensors + Version v1.0 Krea 2 + Size 221.5 MB + Format SafeTensor / unknown precision + SHA256 638be9ada15657f217bfe8049f937a4d75646c63d7dd89be155c19abab1df206 + Published 2026-06-26 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + › "anime style digital painting" + +─────────────────────────────────────────────────────────────────────────────── + 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 6,616 + Rating 100% positive (443/443) + Comments 0 + Published 70 days ago + Download rate 21.9/day since release + + Adoption rank #87 of 150 in this collection (Modest) + Momentum rank #93 of 150 by download rate (Long tail) + + Adoption tier is the percentile of total downloads within this collection; + momentum is the percentile of downloads-per-day since release. Momentum is + ranked rather than measured against a fixed rate, because this base model + is itself new and every LoRA looks fast on an absolute scale. Both numbers + are CivitAI's, read on the scrape date at the foot of this file. + +─────────────────────────────────────────────────────────────────────────────── + CHAINING / STACKING +─────────────────────────────────────────────────────────────────────────────── + Stacks with others not stated + + The author gave no chaining guidance. The rules below are the + general Krea 2 ones, not anything specific to this LoRA. + + General Krea 2 rules that apply here (see REFERENCE.txt): + + • The most important rule is not about chaining but about which + checkpoint you chain onto: the official README says in bold "TRAIN on + Raw and RUN on Turbo", and "We highly recommend using RAW for training + LoRAs and applying them on Turbo for inference." LoRAs trained on Raw + are stated to work well on Turbo. + • Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 + LoRAs are DiT-only, so there is no CLIP side to chain and the text- + encoder branch stays untouched however many LoRAs you add. + • Krea 2 LoRAs stack in practice and the community treats multi-LoRA as + routine - style + character, style + slider, realism + character. + Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one + writes "I recently stacked over 20 of my own LoRAs on a single + generation and the output still came out great." + • Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge + on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is + loaded. The CivitAI 12GB training guide also notes rank affects + composability: "rank 8 produced a lighter style influence that was + easier to stack with other LoRAs, rank 16 provided a useful balance + between concept strength and composability, and rank 32 was better + suited to a precise object or concept intended to take priority over + other stacked LoRAs." + • Accelerator LoRAs are position-sensitive in intent, not in node order: + a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a + sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to- + Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is + explicitly recommended by the 4-step LoRA's author. + • Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI + (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card + says "Keep mu = 1.15. The training targets are anchored to that grid; a + different shift moves the [sampler off them]"; the same reasoning + applies to any Turbo-trained adapter in the chain. + • Reference/edit LoRAs are designed to be stacked under ordinary subject + and style LoRAs - the Identity Edit card states "Composes with your + LoRAs: character/body/style LoRAs stack on top and steer the prior" and + recommends a stacked subject LoRA when unusual subjects drift toward + the base prior. + • Krea 2 LoRAs cannot be chained with LoRAs for any other base model. + Many CivitAI pages carry multi-model titles like "(Krea-2 + ZIT)" or + "[Flux | ZIB | Krea2]" - those are separate files for separate + architectures shipped under one model page, not a combined adapter. + • To combine two full Krea 2 checkpoints rather than adapters, ComfyUI + ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per- + section weights for first., tmlp., txtmlp., tproj., + txtfusion.projector, txtfusion.layerwise_blocks.0-1, + txtfusion.refiner_blocks.0-1, blocks.0-27 and last. + • The author never states which weight family this targets. Silent + degradation, never a hard error, between Raw and Turbo. The two share + one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" + - so every LoRA tensor shape matches on both and ComfyUI/diffuser + +─────────────────────────────────────────────────────────────────────────────── + PROS +─────────────────────────────────────────────────────────────────────────────── + + Widely adopted — 6,616 downloads. [source: CivitAI stats] + + Positively received — 443 thumbs-up with no down-votes. [source: + CivitAI stats] + + Explicit trigger word(s) — anime style digital painting — so the effect + can be turned on and off from the prompt. [source: model metadata] + + Compact at 221 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] + - 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 +─────────────────────────────────────────────────────────────────────────────── + 2.5D anime style for Z-image, Qwen, and Krea 2 + +─────────────────────────────────────────────────────────────────────────────── + 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, 1620x2160) + 02_image.jpeg (image, 1616x2880) + 03_image.jpeg (image, 2048x3072) + 04_image.jpeg (image, 1080x1920) + 05_image.jpeg (image, 1152x832) + 06_image.jpeg (image, 1808x2400) + +─────────────────────────────────────────────────────────────────────────────── + 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 00:16 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/krea2-qwen2511-z-image-illustria-anime-01/k2-illustria3_000000400.safetensors b/krea2-qwen2511-z-image-illustria-anime-01/k2-illustria3_000000400.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..763fded1a938020019c01911fac3eef00fe79217 --- /dev/null +++ b/krea2-qwen2511-z-image-illustria-anime-01/k2-illustria3_000000400.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:638be9ada15657f217bfe8049f937a4d75646c63d7dd89be155c19abab1df206 +size 232234352 diff --git a/krea2-qwen2511-z-image-illustria-anime-01/metadata.json b/krea2-qwen2511-z-image-illustria-anime-01/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..191ab7af10007c8b30c0d49710038cbd9e129fb0 --- /dev/null +++ b/krea2-qwen2511-z-image-illustria-anime-01/metadata.json @@ -0,0 +1,137 @@ +{ + "name": "Krea2+Qwen2511+Z-Image Illustria Anime 01", + "slug": "krea2-qwen2511-z-image-illustria-anime-01", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 2174309, + "version_id": 3072531, + "version_name": "v1.0 Krea 2", + "source": "civitai", + "url": "https://civitai.com/models/2174309", + "repo_id": null, + "file_path": null, + "author": "_Envy_", + "category": { + "key": "style", + "label": "Visual style" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [ + "anime style digital painting" + ], + "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 — 6,616 downloads. [source: CivitAI stats]", + "Positively received — 443 thumbs-up with no down-votes. [source: CivitAI stats]", + "Explicit trigger word(s) — anime style digital painting — so the effect can be turned on and off from the prompt. [source: model metadata]", + "Compact at 221 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]", + "Thin documentation — expect trial and error. [source: description length]" + ], + "file": { + "name": "k2-illustria3_000000400.safetensors", + "size_bytes": 232234352, + "size_mb": 221.48, + "sha256": "638be9ada15657f217bfe8049f937a4d75646c63d7dd89be155c19abab1df206" + }, + "stats": { + "downloads": 6616, + "thumbs_up": 443, + "thumbs_down": 0 + }, + "popularity": { + "downloads": 6616, + "thumbs_up": 443, + "thumbs_down": 0, + "comments": 0, + "approval": "100% positive (443/443)", + "age_days": 70.2, + "downloads_per_day": 21.95, + "rank_by_downloads": 87, + "rank_by_momentum": 93, + "of_total": 150, + "adoption_tier": "Modest", + "momentum": "Long tail" + }, + "selection_score": 11932.0, + "selection_reasons": [ + "6,616 downloads", + "443 thumbs-up", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "1620x2160", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "1616x2880", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "2048x3072", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "1080x1920", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "1152x832", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "1808x2400", + "prompt": "" + } + ], + "tags": [ + "style", + "woman", + "2.5d", + "sexy", + "anime", + "characters" + ], + "scraped_at": "2026-09-05T00:16:12+00:00", + "path": "krea-2/krea2-qwen2511-z-image-illustria-anime-01" +} diff --git a/manga-style-naoki-urasawa/example_images/01_image.jpeg b/manga-style-naoki-urasawa/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..29c79319df11a66a3a6301131991d3163798df2c --- /dev/null +++ b/manga-style-naoki-urasawa/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66032b231607237e278eb5db8be1f507136ced1f602f67a2ad4a662d7f40fd9c +size 2526268 diff --git a/manga-style-naoki-urasawa/example_images/02_image.jpeg b/manga-style-naoki-urasawa/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..650599348327bd8e3f3d1ea787e1fda8f2d40847 --- /dev/null +++ b/manga-style-naoki-urasawa/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e623ee27e89f51c7be0631b08cf9b1b007e41afe352ea68f7486a9cc5c24674e +size 7575577 diff --git a/manga-style-naoki-urasawa/example_images/03_image.jpeg b/manga-style-naoki-urasawa/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..844d8ad603cf4551a58e3a798b47c1eb9fdd01b9 --- /dev/null +++ b/manga-style-naoki-urasawa/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4782cbd75f36bd0b3d5cfda8a954b8116f90a88ef4c8bf6900412f0e00ebaf1a +size 6769826 diff --git a/manga-style-naoki-urasawa/example_images/04_image.jpeg b/manga-style-naoki-urasawa/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..590e8791e54203d35f563a0d81bc3bd048f7e5ce --- /dev/null +++ b/manga-style-naoki-urasawa/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a4245ddb72d9a78f496ad26048198a590e101d917c92889db44898067aa2ca92 +size 6037843 diff --git a/manga-style-naoki-urasawa/example_images/05_image.jpeg b/manga-style-naoki-urasawa/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..5e15b98c7e30d2fe986f79f0444058d1ac3d64a2 --- /dev/null +++ b/manga-style-naoki-urasawa/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bdc83a59ed31392136e3f1d7c4d13d3bbc2014b79d6227b8f6cc427c91dda4e5 +size 4160245 diff --git a/manga-style-naoki-urasawa/example_images/06_image.jpeg b/manga-style-naoki-urasawa/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..18b871bdab503239177c97bd1d3ada556b46da9b --- /dev/null +++ b/manga-style-naoki-urasawa/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:617332bb998dad518b930f199ec225019de7bfa7d5379c44bb233fa3b45ea23d +size 4120100 diff --git a/manga-style-naoki-urasawa/info.txt b/manga-style-naoki-urasawa/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..25cf44df85648db7ee007dc7de045f2c43085a3e --- /dev/null +++ b/manga-style-naoki-urasawa/info.txt @@ -0,0 +1,305 @@ +═══════════════════════════════════════════════════════════════════════════════ + MANGA STYLE (NAOKI URASAWA) +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Visual style + Author Yofaraway + Source https://civitai.com/models/690155 + Provenance CivitAI · model 690155 · version 3087718 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File naoki_urasawa_krea2.safetensors + Version Krea2 v1 + Size 109.0 MB + Format SafeTensor / unknown precision + SHA256 9c21228cb3b1188265ac9951f2f312911fa1977f5367a162921e77e55cd62758 + Published 2026-07-08 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + › "manga style" + +─────────────────────────────────────────────────────────────────────────────── + RECOMMENDED SETTINGS +─────────────────────────────────────────────────────────────────────────────── + Taken verbatim from the author's description: + + LoRA strength 1.3 + Sampler steps 10 + Resolution 944x1408 + CFG / guidance 1 + Sampler Euler simple + +─────────────────────────────────────────────────────────────────────────────── + ADOPTION & POPULARITY +─────────────────────────────────────────────────────────────────────────────── + Downloads 5,438 + Rating 100% positive (485/486) + Comments 0 + Published 58 days ago + Download rate 11.6/day since release + + Adoption rank #106 of 150 in this collection (Modest) + Momentum rank #133 of 150 by download rate (Long tail) + + Adoption tier is the percentile of total downloads within this collection; + momentum is the percentile of downloads-per-day since release. Momentum is + ranked rather than measured against a fixed rate, because this base model + is itself new and every LoRA looks fast on an absolute scale. Both numbers + are CivitAI's, read on the scrape date at the foot of this file. + +─────────────────────────────────────────────────────────────────────────────── + CHAINING / STACKING +─────────────────────────────────────────────────────────────────────────────── + Stacks with others not stated + + The author gave no chaining guidance. The rules below are the + general Krea 2 ones, not anything specific to this LoRA. + + General Krea 2 rules that apply here (see REFERENCE.txt): + + • The most important rule is not about chaining but about which + checkpoint you chain onto: the official README says in bold "TRAIN on + Raw and RUN on Turbo", and "We highly recommend using RAW for training + LoRAs and applying them on Turbo for inference." LoRAs trained on Raw + are stated to work well on Turbo. + • Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 + LoRAs are DiT-only, so there is no CLIP side to chain and the text- + encoder branch stays untouched however many LoRAs you add. + • Krea 2 LoRAs stack in practice and the community treats multi-LoRA as + routine - style + character, style + slider, realism + character. + Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one + writes "I recently stacked over 20 of my own LoRAs on a single + generation and the output still came out great." + • Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge + on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is + loaded. The CivitAI 12GB training guide also notes rank affects + composability: "rank 8 produced a lighter style influence that was + easier to stack with other LoRAs, rank 16 provided a useful balance + between concept strength and composability, and rank 32 was better + suited to a precise object or concept intended to take priority over + other stacked LoRAs." + • Accelerator LoRAs are position-sensitive in intent, not in node order: + a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a + sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to- + Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is + explicitly recommended by the 4-step LoRA's author. + • Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI + (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card + says "Keep mu = 1.15. The training targets are anchored to that grid; a + different shift moves the [sampler off them]"; the same reasoning + applies to any Turbo-trained adapter in the chain. + • Reference/edit LoRAs are designed to be stacked under ordinary subject + and style LoRAs - the Identity Edit card states "Composes with your + LoRAs: character/body/style LoRAs stack on top and steer the prior" and + recommends a stacked subject LoRA when unusual subjects drift toward + the base prior. + • Krea 2 LoRAs cannot be chained with LoRAs for any other base model. + Many CivitAI pages carry multi-model titles like "(Krea-2 + ZIT)" or + "[Flux | ZIB | Krea2]" - those are separate files for separate + architectures shipped under one model page, not a combined adapter. + • To combine two full Krea 2 checkpoints rather than adapters, ComfyUI + ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per- + section weights for first., tmlp., txtmlp., tproj., + txtfusion.projector, txtfusion.layerwise_blocks.0-1, + txtfusion.refiner_blocks.0-1, blocks.0-27 and last. + • The author never states which weight family this targets. Silent + degradation, never a hard error, between Raw and Turbo. The two share + one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" + - so every LoRA tensor shape matches on both and ComfyUI/diffuser + +─────────────────────────────────────────────────────────────────────────────── + PROS +─────────────────────────────────────────────────────────────────────────────── + + Widely adopted — 5,438 downloads. [source: CivitAI stats] + + Positively received — 485 thumbs-up, 1 down. [source: CivitAI stats] + + Explicit trigger word(s) — manga style — so the effect can be turned on + and off from the prompt. [source: model metadata] + + Compact at 109 MB — cheap to stack with other LoRAs. [source: file + size] + + Author published concrete recommended settings. [source: description] + +─────────────────────────────────────────────────────────────────────────────── + CONS & CAVEATS +─────────────────────────────────────────────────────────────────────────────── + - Author never states which generation mode it was trained for; test + against your own workflow before relying on it. [source: description] + +─────────────────────────────────────────────────────────────────────────────── + COMPATIBILITY WARNING +─────────────────────────────────────────────────────────────────────────────── + Silent degradation, never a hard error, between Raw and Turbo. The two + share one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" - so + every LoRA tensor shape matches on both and ComfyUI/diffusers will load an + adapter against either without complaint. The lvladikov 4-step LoRA card + states this for the reverse direction: "Every layer it targets also exists + in Krea 2 Raw, so it will load there without complaint - but that is a side + effect of the shared architecture, not a supported mode", and reports + mixed, subject-dependent results with malformed subjects (duplicated heads, + fused limbs) at low step counts, needing 14+ steps with Raw's CFG on before + output is coherent. The intended direction (train on Raw, run on Turbo) is + officially blessed. The more common practical failure is a settings + mismatch rather than a weight mismatch: running a Turbo-targeted LoRA at + Raw's 52 steps + CFG 3.5, or leaving CFG enabled on Turbo, degrades output + with no warning. Loading a Krea 2 LoRA against a different model family + entirely (Flux, Qwen-Image, Z-Image) is also not an error - ComfyUI logs + "lora key not loaded: {key}" per key and generates as if no LoRA were + applied. + +─────────────────────────────────────────────────────────────────────────────── + AUTHOR'S DESCRIPTION +─────────────────────────────────────────────────────────────────────────────── + KREA 2 Showcase images : + - Turbo model : 10 steps / CFG 1 & Euler simple + + - 944x1408 + + - x2 Hires (10 steps, 0.35 with 4x_NMKD_Siax) + + - Lora strength 1.3 + + IDEOGRAM 4 Json caption style with : + - art_style : Naoki Urasawa style + + - medium : manga + + - aesthetics : black_and_white + + Example caption + KLEIN 9B v1: + Showcase images : + - Distilled model (same prompt images posted with the base model below) + + - 768x1152 + + - Euler / normal / 4 steps / CFG 1 + + - Hires steps : 4 / NMKD_SIAX @0.35 / x2 + + - Lora strength of 0.8-1 + + ZIT v1: + I'm relatively new to Z-Image turbo model, so still experimenting, but I + had good results with these settings : + - 1056*1584 // 10 Steps // CFG 1 + + - DDIM & Beta + + - Lora strength of 0.9 + + I've noticed that using previous prompts (like the ones from Qwen v2) + doesn't work that well, so I've used a vision model (Qwen3-VL-8B) with this + prompt to get updated caption which was giving way better results. + Qwen v2: + Same as for my ghibli lora , I've retrained with everything I've learned so + far, and this version is way better : closer to the target style and more + flexible. + v1: + Trained with Ostris Toolkit: Dataset is 44 images (mainly from "Monster", + "20th Century Boys" and "Pluto" by Naoki Urasawa) with no trigger word (but + sometimes it helps to add "black and white manga style image of"). All the + image have been generated in 4 steps in addition to the lightning lora by + lightx2v The prompt adherence of Qwen is really impressive, but I feel like + the likeness of the style is not as good as with Flux (even tho is still + good), but I'm still experimenting so it's just an early opinion. + Flux Trained on images from the manga "Monster" and "20th Century boys" by + Naoki Urasawa. + Start the prompt with : + A naoUrasawa style image of Using this upscaler seems to improve the style + towards a true manga scan most of the time (and get rid of colors). I've + used it with hires fix and denoising strength of 0.5. + +─────────────────────────────────────────────────────────────────────────────── + 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, 1120x1664) + 02_image.jpeg (image, 1888x2816) + 03_image.jpeg (image, 1888x2816) + 04_image.jpeg (image, 1888x2816) + 05_image.jpeg (image, 1280x1920) + 06_image.jpeg (image, 1280x1920) + +─────────────────────────────────────────────────────────────────────────────── + 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 00:16 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/manga-style-naoki-urasawa/metadata.json b/manga-style-naoki-urasawa/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..f754435b2f778b101091ac97bbafe6e463a21dee --- /dev/null +++ b/manga-style-naoki-urasawa/metadata.json @@ -0,0 +1,147 @@ +{ + "name": "Manga style (Naoki Urasawa)", + "slug": "manga-style-naoki-urasawa", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 690155, + "version_id": 3087718, + "version_name": "Krea2 v1", + "source": "civitai", + "url": "https://civitai.com/models/690155", + "repo_id": null, + "file_path": null, + "author": "Yofaraway", + "category": { + "key": "style", + "label": "Visual style" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [ + "manga style" + ], + "recommended_settings": { + "LoRA strength": "1.3", + "Sampler steps": "10", + "Resolution": "944x1408", + "CFG / guidance": "1", + "Sampler": "Euler simple" + }, + "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 — 5,438 downloads. [source: CivitAI stats]", + "Positively received — 485 thumbs-up, 1 down. [source: CivitAI stats]", + "Explicit trigger word(s) — manga style — so the effect can be turned on and off from the prompt. [source: model metadata]", + "Compact at 109 MB — cheap to stack with other LoRAs. [source: file size]", + "Author published concrete recommended settings. [source: description]" + ], + "cons": [ + "Author never states which generation mode it was trained for; test against your own workflow before relying on it. [source: description]" + ], + "file": { + "name": "naoki_urasawa_krea2.safetensors", + "size_bytes": 114326360, + "size_mb": 109.03, + "sha256": "9c21228cb3b1188265ac9951f2f312911fa1977f5367a162921e77e55cd62758" + }, + "stats": { + "downloads": 5438, + "thumbs_up": 485, + "thumbs_down": 1 + }, + "popularity": { + "downloads": 5438, + "thumbs_up": 485, + "thumbs_down": 1, + "comments": 0, + "approval": "100% positive (485/486)", + "age_days": 58.1, + "downloads_per_day": 11.58, + "rank_by_downloads": 106, + "rank_by_momentum": 133, + "of_total": 150, + "adoption_tier": "Modest", + "momentum": "Long tail" + }, + "selection_score": 11658.0, + "selection_reasons": [ + "5,438 downloads", + "485 thumbs-up", + "documented triggers + usage notes", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "1120x1664", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "1888x2816", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "1888x2816", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "1888x2816", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "1280x1920", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "1280x1920", + "prompt": "" + } + ], + "tags": [ + "manga", + "drawing", + "monster", + "naoki urasawa", + "artist", + "20th century boys", + "style", + "pluto", + "illustration" + ], + "scraped_at": "2026-09-05T00:16:11+00:00", + "path": "krea-2/manga-style-naoki-urasawa" +} diff --git a/manga-style-naoki-urasawa/naoki_urasawa_krea2.safetensors b/manga-style-naoki-urasawa/naoki_urasawa_krea2.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..ba3498056e6d953edf198413739fc602a9d8f35d --- /dev/null +++ b/manga-style-naoki-urasawa/naoki_urasawa_krea2.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c21228cb3b1188265ac9951f2f312911fa1977f5367a162921e77e55cd62758 +size 114326360 diff --git a/minimalist-vector-art/MinimalisticVectorArtKrea2.safetensors b/minimalist-vector-art/MinimalisticVectorArtKrea2.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..9188c63c606626ad96ed995d71e95dd923c47dcf --- /dev/null +++ b/minimalist-vector-art/MinimalisticVectorArtKrea2.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a63836304fb35eecc6fd70b8752f6e5d73dce8852256bd98f4ec7ae679a7157 +size 228587744 diff --git a/minimalist-vector-art/example_images/01_image.jpeg b/minimalist-vector-art/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..7a3cdd7e4af871802ea906c9c98215da172b9df6 --- /dev/null +++ b/minimalist-vector-art/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:61f9a63fbcee0771012330d5c820ecabcacf4a53db708d4fe281df70d02362f6 +size 915664 diff --git a/minimalist-vector-art/example_images/02_image.jpeg b/minimalist-vector-art/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..543042b75d7dc7718e093917cf5b9d8e2ebdc0f1 --- /dev/null +++ b/minimalist-vector-art/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:950d5b1f0aff7451bdd1378216e66b84302f6e37674339ccdb28a214d0ee94a6 +size 1037482 diff --git a/minimalist-vector-art/example_images/03_image.jpeg b/minimalist-vector-art/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..d71d988f5934ac49dfdb387016ea1d267d21534d --- /dev/null +++ b/minimalist-vector-art/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1903c9b41b6c36384f226c4215373f2cd280b166408fca1bde79badf2e8ee452 +size 615186 diff --git a/minimalist-vector-art/example_images/04_image.jpeg b/minimalist-vector-art/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..e134d542a4719c7fffcf9efeb5ff9118eb115a9c --- /dev/null +++ b/minimalist-vector-art/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4c9d3b73c88f051a3effcd4cc23c13e31de8844bae3dad6d941ba9c78833ca9e +size 851823 diff --git a/minimalist-vector-art/example_images/05_image.jpeg b/minimalist-vector-art/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..68b41a6e1f394a00fde0f4b6114dbf490fb34416 --- /dev/null +++ b/minimalist-vector-art/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:44f9e7ba7b9f6848ea3d97edadc892a7da9251ab3117e8899195f2f00a6697df +size 612681 diff --git a/minimalist-vector-art/example_images/06_image.jpeg b/minimalist-vector-art/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..5067437467daf33b6a6931b3c272e5990d1d68c5 --- /dev/null +++ b/minimalist-vector-art/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5bca1dc355d6e600f8c24e577bd7caa162bbb2b7eee556f7bf4dc5d7729db194 +size 822655 diff --git a/minimalist-vector-art/info.txt b/minimalist-vector-art/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..e149d478494b28a5d5898ddea37a106a72a0601f --- /dev/null +++ b/minimalist-vector-art/info.txt @@ -0,0 +1,283 @@ +═══════════════════════════════════════════════════════════════════════════════ + MINIMALIST VECTOR ART +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + 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/658816 + Provenance CivitAI · model 658816 · version 3177745 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File MinimalisticVectorArtKrea2.safetensors + Version Krea2 + Size 218.0 MB + Format SafeTensor / unknown precision + SHA256 9a63836304fb35eecc6fd70b8752f6e5d73dce8852256bd98f4ec7ae679a7157 + Published 2026-08-24 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + › "Minimalist Vector Art, ArsMJStyle," + +─────────────────────────────────────────────────────────────────────────────── + 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 5,427 + Rating 100% positive (533/533) + Comments 2 + Published 12 days ago + Download rate 29.5/day since release + + Adoption rank #107 of 150 in this collection (Modest) + Momentum rank #76 of 150 by download rate (Long tail · fresh release) + + Adoption tier is the percentile of total downloads within this collection; + momentum is the percentile of downloads-per-day since release. Momentum is + ranked rather than measured against a fixed rate, because this base model + is itself new and every LoRA looks fast on an absolute scale. Both numbers + are CivitAI's, read on the scrape date at the foot of this file. + +─────────────────────────────────────────────────────────────────────────────── + CHAINING / STACKING +─────────────────────────────────────────────────────────────────────────────── + Stacks with others not stated + + The author gave no chaining guidance. The rules below are the + general Krea 2 ones, not anything specific to this LoRA. + + General Krea 2 rules that apply here (see REFERENCE.txt): + + • The most important rule is not about chaining but about which + checkpoint you chain onto: the official README says in bold "TRAIN on + Raw and RUN on Turbo", and "We highly recommend using RAW for training + LoRAs and applying them on Turbo for inference." LoRAs trained on Raw + are stated to work well on Turbo. + • Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 + LoRAs are DiT-only, so there is no CLIP side to chain and the text- + encoder branch stays untouched however many LoRAs you add. + • Krea 2 LoRAs stack in practice and the community treats multi-LoRA as + routine - style + character, style + slider, realism + character. + Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one + writes "I recently stacked over 20 of my own LoRAs on a single + generation and the output still came out great." + • Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge + on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is + loaded. The CivitAI 12GB training guide also notes rank affects + composability: "rank 8 produced a lighter style influence that was + easier to stack with other LoRAs, rank 16 provided a useful balance + between concept strength and composability, and rank 32 was better + suited to a precise object or concept intended to take priority over + other stacked LoRAs." + • Accelerator LoRAs are position-sensitive in intent, not in node order: + a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a + sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to- + Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is + explicitly recommended by the 4-step LoRA's author. + • Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI + (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card + says "Keep mu = 1.15. The training targets are anchored to that grid; a + different shift moves the [sampler off them]"; the same reasoning + applies to any Turbo-trained adapter in the chain. + • Reference/edit LoRAs are designed to be stacked under ordinary subject + and style LoRAs - the Identity Edit card states "Composes with your + LoRAs: character/body/style LoRAs stack on top and steer the prior" and + recommends a stacked subject LoRA when unusual subjects drift toward + the base prior. + • Krea 2 LoRAs cannot be chained with LoRAs for any other base model. + Many CivitAI pages carry multi-model titles like "(Krea-2 + ZIT)" or + "[Flux | ZIB | Krea2]" - those are separate files for separate + architectures shipped under one model page, not a combined adapter. + • To combine two full Krea 2 checkpoints rather than adapters, ComfyUI + ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per- + section weights for first., tmlp., txtmlp., tproj., + txtfusion.projector, txtfusion.layerwise_blocks.0-1, + txtfusion.refiner_blocks.0-1, blocks.0-27 and last. + • The author never states which weight family this targets. Silent + degradation, never a hard error, between Raw and Turbo. The two share + one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" + - so every LoRA tensor shape matches on both and ComfyUI/diffuser + +─────────────────────────────────────────────────────────────────────────────── + PROS +─────────────────────────────────────────────────────────────────────────────── + + Widely adopted — 5,427 downloads. [source: CivitAI stats] + + Positively received — 533 thumbs-up with no down-votes. [source: + CivitAI stats] + + Explicit trigger word(s) — Minimalist Vector Art, ArsMJStyle, — 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 +─────────────────────────────────────────────────────────────────────────────── + The Lora is trained on MidJourney Images that use my "Personalised Style" + + Minimalist Flat Vector Art with eastern themes. + + Тrigger words are ArsMJStyle, Minimalist Vector Art + If you like the Monochrome showcases you can enchance the effect by adding: + Silhouette, monochrome, greyscale, simple background + + Works great in the 0.4 - 2 range Gives the effects I was aiming for in the + 1.2 - 1.5 range. You don't have to keep the prompt simplistic in this + range. Start around 1.2 and see what works for you :) + +─────────────────────────────────────────────────────────────────────────────── + EXAMPLE MEDIA +─────────────────────────────────────────────────────────────────────────────── + 6 preview(s) in ./example_images/ — community results for + this LoRA. Videos keep a matching _poster.jpg still frame. + + No prompts are recorded below: CivitAI's public API flags these posts as + having generation metadata but does not return it, so the prompts cannot be + scraped. Open the model page to read them. + + 01_image.jpeg (image, 1024x1024) + 02_image.jpeg (image, 1024x1024) + 03_image.jpeg (image, 1024x1024) + 04_image.jpeg (image, 1024x1024) + 05_image.jpeg (image, 1024x1024) + 06_image.jpeg (image, 1024x1024) + +─────────────────────────────────────────────────────────────────────────────── + HOW TO USE (COMFYUI) +─────────────────────────────────────────────────────────────────────────────── + 1. Update ComfyUI to at least v0.26.0 (released 2026-06-23), which added + local Krea2 support. For reference-image / style-reference / edit LoRAs you + need at least v0.29.0 (2026-07-29), which added regular and timestep-zero + reference images for Krea 2. ModelMergeKrea2 arrived in v0.27.0 + (2026-06-30). + + 2. Put the base stack in place (files from huggingface.co/Comfy-Org/Krea-2, + which is not gated): + - ComfyUI/models/diffusion_models/krea2_turbo_fp8_scaled.safetensors + (documented recommended default; the repo also holds krea2_turbo_bf16, + _int8_convrot, _mxfp8, _nvfp4 and krea2_raw_bf16 / _fp8_scaled / + _int8_convrot) + - ComfyUI/models/text_encoders/qwen3vl_4b_fp8_scaled.safetensors (bf16 + also available) + - ComfyUI/models/vae/qwen_image_vae.safetensors + + 3. Put the LoRA .safetensors in ComfyUI/models/loras/ - flat, no per-model + subfolder required. + + 4. Build or load the graph. The official image_krea2_turbo_t2i.json + template is: UNETLoader (krea2_turbo_fp8_scaled.safetensors) -> + LoraLoaderModelOnly -> KSampler; CLIPLoader + (qwen3vl_4b_fp8_scaled.safetensors, type "krea2") -> text encode -> + KSampler positive, with ConditioningZeroOut into negative; VAELoader + (qwen_image_vae.safetensors) -> VAEDecode -> SaveImage; EmptyLatentImage + 1024x1024 driven by ResolutionSelector. An int8 sibling template + (image_krea2_turbo_t2i_int8.json) also ships. + + 5. Load the LoRA with LoraLoaderModelOnly, not LoraLoader - both official + templates do. Krea 2 LoRAs patch the DiT only; note that the DiT's own + text_fusion / txtfusion blocks are a normal LoRA target (the official style + LoRAs patch them), but the Qwen3-VL text encoder itself is not. Set + strength 1.0 for Krea's official style LoRAs (the shipped template uses + 0.8) and add the LoRA's trigger word to the prompt. + + 6. Sampler settings: steps 8, cfg 1, sampler euler, scheduler simple, + denoise 1.0. Do not add a ModelSampling node for plain t2i - ComfyUI + already applies shift 1.15 from Krea2.sampling_settings. + + 7. For style reference, use the dedicated template: UNETLoader + krea2_turbo_int8_convrot.safetensors, LoraLoaderModelOnly + krea2_style_reference.safetensors at 1.0, ModelSamplingFlux (1.15, 0.5), + TextEncodeQwenImageEditPlus fed by LoadImage, + FluxKontextMultiReferenceLatentMethod set to index_timestep_zero, and + SamplerCustomAdvanced with KSamplerSelect euler + BasicScheduler simple/8 + + CFGGuider cfg 1. + + 8. For instruction editing (Krea 2 Identity Edit), stock nodes are not + enough: install the ComfyUI-Krea2Edit node pack, which ships its own + workflows, and run Turbo at 8-12 steps / cfg 1 (or Raw at 20 steps / cfg 3 + for removals) with LoRA strength 1.0. + + 9. Low-VRAM alternative: install city96's ComfyUI-GGUF and use a community + GGUF build (realrebelai/KREA-2_GGUFs, quantized from Turbo; + molbal/krea2-gguf, quantized from Raw) with the Unet Loader (GGUF) node in + place of UNETLoader. There is no official GGUF release. LoRA loading is + unchanged. + +─────────────────────────────────────────────────────────────────────────────── + Scraped by lorakit on 2026-09-05 00:15 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/minimalist-vector-art/metadata.json b/minimalist-vector-art/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..30539d4e12d422d1a82852659a2eee26d75a4003 --- /dev/null +++ b/minimalist-vector-art/metadata.json @@ -0,0 +1,138 @@ +{ + "name": "Minimalist vector art", + "slug": "minimalist-vector-art", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 658816, + "version_id": 3177745, + "version_name": "Krea2", + "source": "civitai", + "url": "https://civitai.com/models/658816", + "repo_id": null, + "file_path": null, + "author": "ArsMachina", + "category": { + "key": "style", + "label": "Visual style" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [ + "Minimalist Vector Art, ArsMJStyle," + ], + "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 — 5,427 downloads. [source: CivitAI stats]", + "Positively received — 533 thumbs-up with no down-votes. [source: CivitAI stats]", + "Explicit trigger word(s) — Minimalist Vector Art, ArsMJStyle, — 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": "MinimalisticVectorArtKrea2.safetensors", + "size_bytes": 228587744, + "size_mb": 218.0, + "sha256": "9a63836304fb35eecc6fd70b8752f6e5d73dce8852256bd98f4ec7ae679a7157" + }, + "stats": { + "downloads": 5427, + "thumbs_up": 533, + "thumbs_down": 0 + }, + "popularity": { + "downloads": 5427, + "thumbs_up": 533, + "thumbs_down": 0, + "comments": 2, + "approval": "100% positive (533/533)", + "age_days": 11.7, + "downloads_per_day": 29.46, + "rank_by_downloads": 107, + "rank_by_momentum": 76, + "of_total": 150, + "adoption_tier": "Modest", + "momentum": "Long tail · fresh release" + }, + "selection_score": 12223.0, + "selection_reasons": [ + "5,427 downloads", + "533 thumbs-up", + "documented triggers + usage notes", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "1024x1024", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "1024x1024", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "1024x1024", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "1024x1024", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "1024x1024", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "1024x1024", + "prompt": "" + } + ], + "tags": [ + "style", + "midjourney", + "vectorart", + "silhouette", + "midjourney v6", + "arsmjstyle", + "minimalist" + ], + "scraped_at": "2026-09-05T00:15:50+00:00", + "path": "krea-2/minimalist-vector-art" +} diff --git a/purelens-realism-ideogram-4-krea-2/example_images/01_image.jpeg b/purelens-realism-ideogram-4-krea-2/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..759c07090dcc4f84dc29cc612030b3a83be31c8f --- /dev/null +++ b/purelens-realism-ideogram-4-krea-2/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e1be6fd24a47e5afdc05397abaa80db633369773634c5ebc80fbd8dbf7f8542 +size 2411201 diff --git a/purelens-realism-ideogram-4-krea-2/example_images/02_image.jpeg b/purelens-realism-ideogram-4-krea-2/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..c566287a67ba5fb363eb72034090a8caa835e880 --- /dev/null +++ b/purelens-realism-ideogram-4-krea-2/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ae604232e366704140e1b155225f232c2169b39528a1fd28560d0f304e041c7 +size 2451835 diff --git a/purelens-realism-ideogram-4-krea-2/example_images/03_image.jpeg b/purelens-realism-ideogram-4-krea-2/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..e6239fa7d5eb76bae3649a7b31131d3140dce527 --- /dev/null +++ b/purelens-realism-ideogram-4-krea-2/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:24bc196bc12bc8232509630dbe23216d677a680ffc4462a3c55252aac2da47a1 +size 1721190 diff --git a/purelens-realism-ideogram-4-krea-2/example_images/04_image.jpeg b/purelens-realism-ideogram-4-krea-2/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..9cfbe3293cf3349aa6cc3b2661f9f576dd3fe87c --- /dev/null +++ b/purelens-realism-ideogram-4-krea-2/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:62d09d3b2a1a2fe69656b54795f3cd573475a474df0524c354125f28156d6ff9 +size 962916 diff --git a/purelens-realism-ideogram-4-krea-2/example_images/05_image.jpeg b/purelens-realism-ideogram-4-krea-2/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..de60a7c087d6d8ee34af0d65259380bb5df0c291 --- /dev/null +++ b/purelens-realism-ideogram-4-krea-2/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3efa851a91443506abdb4d5377bceedc798ca2b018d12455f0a3d59493984d61 +size 3553341 diff --git a/purelens-realism-ideogram-4-krea-2/example_images/06_image.jpeg b/purelens-realism-ideogram-4-krea-2/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..fe7c5517d25030f88038042fcb8f024e781a710b --- /dev/null +++ b/purelens-realism-ideogram-4-krea-2/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2e659d0b9cd6bad75d05828d628b9809093be2970829451b54718f2e49bd3f6c +size 2529442 diff --git a/purelens-realism-ideogram-4-krea-2/info.txt b/purelens-realism-ideogram-4-krea-2/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..709134951d2c83b38cd210b469ea394aaadebd65 --- /dev/null +++ b/purelens-realism-ideogram-4-krea-2/info.txt @@ -0,0 +1,260 @@ +═══════════════════════════════════════════════════════════════════════════════ + PURELENS REALISM (IDEOGRAM 4-KREA 2) +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Photorealism + Author goldennyks76 + Source https://civitai.com/models/2728005 + Provenance CivitAI · model 2728005 · version 3114242 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File purelens_krea2.safetensors + Version Krea 2 v1.0 + Size 327.0 MB + Format SafeTensor / unknown precision + SHA256 e4948147a904a480a46a0beb81a9fc579043e7d670bc190ab827d0dee3cb80fe + Published 2026-07-09 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + › "purelens" + +─────────────────────────────────────────────────────────────────────────────── + RECOMMENDED SETTINGS +─────────────────────────────────────────────────────────────────────────────── + Taken verbatim from the author's description: + + LoRA strength 0.6–1.5 + +─────────────────────────────────────────────────────────────────────────────── + ADOPTION & POPULARITY +─────────────────────────────────────────────────────────────────────────────── + Downloads 7,657 + Rating 100% positive (344/344) + Comments 15 + Published 57 days ago + Download rate 118.6/day since release + + Adoption rank #76 of 150 in this collection (Modest) + Momentum rank #21 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 — 7,657 downloads. [source: CivitAI stats] + + Positively received — 344 thumbs-up with no down-votes. [source: + CivitAI stats] + + Explicit trigger word(s) — purelens — so the effect can be turned on + and off from the prompt. [source: model metadata] + + Compact at 327 MB — cheap to stack with other LoRAs. [source: file + size] + + Author published concrete recommended settings. [source: description] + +─────────────────────────────────────────────────────────────────────────────── + CONS & CAVEATS +─────────────────────────────────────────────────────────────────────────────── + - Author never states which generation mode it was trained for; test + against your own workflow before relying on it. [source: description] + +─────────────────────────────────────────────────────────────────────────────── + COMPATIBILITY WARNING +─────────────────────────────────────────────────────────────────────────────── + Silent degradation, never a hard error, between Raw and Turbo. The two + share one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" - so + every LoRA tensor shape matches on both and ComfyUI/diffusers will load an + adapter against either without complaint. The lvladikov 4-step LoRA card + states this for the reverse direction: "Every layer it targets also exists + in Krea 2 Raw, so it will load there without complaint - but that is a side + effect of the shared architecture, not a supported mode", and reports + mixed, subject-dependent results with malformed subjects (duplicated heads, + fused limbs) at low step counts, needing 14+ steps with Raw's CFG on before + output is coherent. The intended direction (train on Raw, run on Turbo) is + officially blessed. The more common practical failure is a settings + mismatch rather than a weight mismatch: running a Turbo-targeted LoRA at + Raw's 52 steps + CFG 3.5, or leaving CFG enabled on Turbo, degrades output + with no warning. Loading a Krea 2 LoRA against a different model family + entirely (Flux, Qwen-Image, Z-Image) is also not an error - ComfyUI logs + "lora key not loaded: {key}" per key and generates as if no LoRA were + applied. + +─────────────────────────────────────────────────────────────────────────────── + AUTHOR'S DESCRIPTION +─────────────────────────────────────────────────────────────────────────────── + Pure Lens is a realism LoRA trained on high-quality photographs and images + converted from raw formats (DNG, ARW, CR2, NEF...). The dataset consists of + natural, sharp images — not studio shoots. Clean, crisp visuals that bring + out texture, pores, and fine details. Works well with other LoRAs. + + 🔵 Krea v1.0 — Latest This is the Krea version of previous model. Compared + to the Ideogram version, this dataset is significantly broader and features + much better captioning. + Strength: 0.6 - 1.5 + + 🟢 Ideogram v1.0 Refined version. Consistent results, improved skin texture + and detail quality. Works great with portraits, close-ups, and body shots. + Now also performs better on medium shots and long shots. + + 🟡 Ideogram v0.5 Beta version trained on a small dataset. Adds realistic + skin texture and clean details. Works best with portraits, close-ups, and + body shots. May be inconsistent on complex scenes. + +─────────────────────────────────────────────────────────────────────────────── + 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, 1152x1536) + 02_image.jpeg (image, 1152x1536) + 03_image.jpeg (image, 1024x1440) + 04_image.jpeg (image, 928x1152) + 05_image.jpeg (image, 1256x1672) + 06_image.jpeg (image, 1256x1672) + +─────────────────────────────────────────────────────────────────────────────── + 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 00:15 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/purelens-realism-ideogram-4-krea-2/metadata.json b/purelens-realism-ideogram-4-krea-2/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..c1b99d1c86ab3f362557e7afc34b4a2c58606983 --- /dev/null +++ b/purelens-realism-ideogram-4-krea-2/metadata.json @@ -0,0 +1,140 @@ +{ + "name": "PureLens Realism (Ideogram 4-Krea 2)", + "slug": "purelens-realism-ideogram-4-krea-2", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 2728005, + "version_id": 3114242, + "version_name": "Krea 2 v1.0", + "source": "civitai", + "url": "https://civitai.com/models/2728005", + "repo_id": null, + "file_path": null, + "author": "goldennyks76", + "category": { + "key": "photoreal", + "label": "Photorealism" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [ + "purelens" + ], + "recommended_settings": { + "LoRA strength": "0.6–1.5" + }, + "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 — 7,657 downloads. [source: CivitAI stats]", + "Positively received — 344 thumbs-up with no down-votes. [source: CivitAI stats]", + "Explicit trigger word(s) — purelens — so the effect can be turned on and off from the prompt. [source: model metadata]", + "Compact at 327 MB — cheap to stack with other LoRAs. [source: file size]", + "Author published concrete recommended settings. [source: description]" + ], + "cons": [ + "Author never states which generation mode it was trained for; test against your own workflow before relying on it. [source: description]" + ], + "file": { + "name": "purelens_krea2.safetensors", + "size_bytes": 342849952, + "size_mb": 326.97, + "sha256": "e4948147a904a480a46a0beb81a9fc579043e7d670bc190ab827d0dee3cb80fe" + }, + "stats": { + "downloads": 7657, + "thumbs_up": 344, + "thumbs_down": 0 + }, + "popularity": { + "downloads": 7657, + "thumbs_up": 344, + "thumbs_down": 0, + "comments": 15, + "approval": "100% positive (344/344)", + "age_days": 57.2, + "downloads_per_day": 118.62, + "rank_by_downloads": 76, + "rank_by_momentum": 21, + "of_total": 150, + "adoption_tier": "Modest", + "momentum": "Trending" + }, + "selection_score": 12185.0, + "selection_reasons": [ + "7,657 downloads", + "344 thumbs-up", + "documented triggers + usage notes", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "1152x1536", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "1152x1536", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "1024x1440", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "928x1152", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "1256x1672", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "1256x1672", + "prompt": "" + } + ], + "tags": [ + "ideogram", + "woman", + "skin", + "photorealistic", + "raw", + "style" + ], + "scraped_at": "2026-09-05T00:15:48+00:00", + "path": "krea-2/purelens-realism-ideogram-4-krea-2" +} diff --git a/purelens-realism-ideogram-4-krea-2/purelens_krea2.safetensors b/purelens-realism-ideogram-4-krea-2/purelens_krea2.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..5432c20f9650827935b2834dc9240b3cf8afa731 --- /dev/null +++ b/purelens-realism-ideogram-4-krea-2/purelens_krea2.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e4948147a904a480a46a0beb81a9fc579043e7d670bc190ab827d0dee3cb80fe +size 342849952 diff --git a/ray-krea2-style-loras/Ray Versatile Face v1.safetensors b/ray-krea2-style-loras/Ray Versatile Face v1.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..cda6987f68e7b5f863872304b059a4e4eecdeaf8 --- /dev/null +++ b/ray-krea2-style-loras/Ray Versatile Face v1.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:89debcdae181477684af446c8208a39a40612ba2dd4e3e6f53808ecd093e85d2 +size 234702984 diff --git a/ray-krea2-style-loras/example_images/01_image.jpeg b/ray-krea2-style-loras/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..d098b9a09d427056b3692200579a54d9eccb1fbd --- /dev/null +++ b/ray-krea2-style-loras/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f5b883fdf31c09cbe941138c4a2775de0fca42a9dd5804ee8f3430aa557471ab +size 2507433 diff --git a/ray-krea2-style-loras/example_images/02_image.jpeg b/ray-krea2-style-loras/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..faee5c7c312af82dfd17911e9fa7b6b4a8dbaa56 --- /dev/null +++ b/ray-krea2-style-loras/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a1442d0514f01b34b069398ff69184166aa968e407da1995bf468ef2cbf3dd57 +size 2532350 diff --git a/ray-krea2-style-loras/example_images/03_image.jpeg b/ray-krea2-style-loras/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..79a67e0d4c308a81643c613263136f5bdba893e7 --- /dev/null +++ b/ray-krea2-style-loras/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f8fb3b5f0925d98b8d24d25910e89b162ace15eca261cfe541394aa941a3784 +size 2160222 diff --git a/ray-krea2-style-loras/example_images/04_image.jpeg b/ray-krea2-style-loras/example_images/04_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..78b7fbaa7afcae3bc7ace9985de62c05cb3ef3de --- /dev/null +++ b/ray-krea2-style-loras/example_images/04_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0410a9d92e79761ad6855a3c52cca7cac00124758809e6ab2381904ecb731c1f +size 1935057 diff --git a/ray-krea2-style-loras/example_images/05_image.jpeg b/ray-krea2-style-loras/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..a3f50a74614408cf1ab679d0a7b5a28c46c2a550 --- /dev/null +++ b/ray-krea2-style-loras/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:681f76bba6a65330af085dc407070519eacb42e0d13d32bfbe05ec8ed1a4a2d1 +size 2108571 diff --git a/ray-krea2-style-loras/example_images/06_image.jpeg b/ray-krea2-style-loras/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..c3d88f264802bed33e820712d1d3fe3c96ea098b --- /dev/null +++ b/ray-krea2-style-loras/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b227a5af73337ada0f5aa5ad6c3206856d204b2c758759b73a3f03363328bcc +size 2598030 diff --git a/ray-krea2-style-loras/info.txt b/ray-krea2-style-loras/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..2ee60d2ab36a3c249efad9b86f8ff4e6030f6701 --- /dev/null +++ b/ray-krea2-style-loras/info.txt @@ -0,0 +1,339 @@ +═══════════════════════════════════════════════════════════════════════════════ + RAY KREA2 STYLE LORAS +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode not declared by the author + Checkpoint not determined — verify before use + Category Detail / anatomy fixer + Author Ray3780 + Source https://civitai.com/models/2802215 + Provenance CivitAI · model 2802215 · version 3280920 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File Ray Versatile Face v1.safetensors + Version Ray Versatile Face + Size 223.8 MB + Format SafeTensor / unknown precision + SHA256 89debcdae181477684af446c8208a39a40612ba2dd4e3e6f53808ecd093e85d2 + Published 2026-08-31 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + No trigger word. The LoRA applies as soon as it is loaded — + strength is the only way to dial the effect up or down. + +─────────────────────────────────────────────────────────────────────────────── + RECOMMENDED SETTINGS +─────────────────────────────────────────────────────────────────────────────── + The author did not publish specific settings. + No author settings were published. Community baseline: steps Turbo: 8 + (official README and HF card CLI example; the ComfyUI template ships 8). + Raw: 52 (official README). The krea-2 CLI's generic default is 28. + Community Turbo LoRAs commonly suggest 8-12 (e.g. the Identity Edit card + recommends Turbo at 8-12, or Raw at 20 for removals). A community 4-step + LoRA on Turbo makes 4 valid; its author calls 2 steps an out-of-spec + preview mode., cfg Turbo: CFG disabled - guidance_scale 0.0 in diffusers + and --cfg 0.0 in the krea-2 CLI, which is cfg 1.0 in ComfyUI's KSampler + (the official template ships cfg 1). These are the same thing; do not set + ComfyUI cfg to 0. Raw: CFG 3.5 (official README; CLI generic default 4.5). + Negative prompting on Turbo is effectively unused - the official templates + route ConditioningZeroOut into the negative slot., sampler ComfyUI: + KSampler with sampler euler, scheduler simple, denoise 1.0 (verified in + image_krea2_turbo_t2i.json: [seed, 'randomize', 8, 1, 'euler', 'simple', + 1]). Timestep shift mu = 1.15 for Turbo; ComfyUI bakes this in via + Krea2.sampling_settings = {"multiplier": 1.0, "shift": 1.15}, so the plain + t2i template needs no ModelSampling node, while the style-reference + template adds ModelSamplingFlux at max_shift 1.15 / base_shift 0.5. The CLI + exposes --mu 1.15 for Turbo; Raw uses a resolution-derived mu interpolated + between --y1 0.5 and --y2 1.15 (ai-toolkit mirrors this as base_shift 0.5 / + max_shift 1.15 over image_seq_len 256-6400, exponential time shift)., + resolution 1024x1024 default. Turbo: 1K-2K (the official CLI example + generates 2048x2048; ComfyUI's ResolutionSelector defaults to 1 megapixel, + set 2.0 for 2K). Raw: "trained to generate upto 1k resolution". + Width/height are padded up to a multiple of 16., lora_strength 1.0 for + Krea's nine official style LoRAs (docs.comfy.org lists Recommended Strength + 1.0 for every one), though the shipped t2i template's LoraLoaderModelOnly + is set to 0.8. The style-reference LoRA is used at 1.0 in its template. The + 4-step distill LoRA documents 1.0 as its trained point, 1.0-1.5 as coherent + extrapolation and above ~1.5 as speckle artefacting. Community subject + LoRAs are typically recommended around 0.8-1.2 and dropped toward 0.6-0.9 + when several are stacked.. + +─────────────────────────────────────────────────────────────────────────────── + ADOPTION & POPULARITY +─────────────────────────────────────────────────────────────────────────────── + Downloads 7,530 + Rating 100% positive (337/337) + Comments 17 + Published 5 days ago + Download rate 66.9/day since release + + Adoption rank #77 of 150 in this collection (Modest) + Momentum rank #41 of 150 by download rate (Steady · fresh release) + + Adoption tier is the percentile of total downloads within this collection; + momentum is the percentile of downloads-per-day since release. Momentum is + ranked rather than measured against a fixed rate, because this base model + is itself new and every LoRA looks fast on an absolute scale. Both numbers + are CivitAI's, read on the scrape date at the foot of this file. + +─────────────────────────────────────────────────────────────────────────────── + CHAINING / STACKING +─────────────────────────────────────────────────────────────────────────────── + Stacks with others not stated + + The author gave no chaining guidance. The rules below are the + general Krea 2 ones, not anything specific to this LoRA. + + General Krea 2 rules that apply here (see REFERENCE.txt): + + • The most important rule is not about chaining but about which + checkpoint you chain onto: the official README says in bold "TRAIN on + Raw and RUN on Turbo", and "We highly recommend using RAW for training + LoRAs and applying them on Turbo for inference." LoRAs trained on Raw + are stated to work well on Turbo. + • Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 + LoRAs are DiT-only, so there is no CLIP side to chain and the text- + encoder branch stays untouched however many LoRAs you add. + • Krea 2 LoRAs stack in practice and the community treats multi-LoRA as + routine - style + character, style + slider, realism + character. + Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one + writes "I recently stacked over 20 of my own LoRAs on a single + generation and the output still came out great." + • Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge + on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is + loaded. The CivitAI 12GB training guide also notes rank affects + composability: "rank 8 produced a lighter style influence that was + easier to stack with other LoRAs, rank 16 provided a useful balance + between concept strength and composability, and rank 32 was better + suited to a precise object or concept intended to take priority over + other stacked LoRAs." + • Accelerator LoRAs are position-sensitive in intent, not in node order: + a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a + sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to- + Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is + explicitly recommended by the 4-step LoRA's author. + • Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI + (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card + says "Keep mu = 1.15. The training targets are anchored to that grid; a + different shift moves the [sampler off them]"; the same reasoning + applies to any Turbo-trained adapter in the chain. + • Reference/edit LoRAs are designed to be stacked under ordinary subject + and style LoRAs - the Identity Edit card states "Composes with your + LoRAs: character/body/style LoRAs stack on top and steer the prior" and + recommends a stacked subject LoRA when unusual subjects drift toward + the base prior. + • Krea 2 LoRAs cannot be chained with LoRAs for any other base model. + Many CivitAI pages carry multi-model titles like "(Krea-2 + ZIT)" or + "[Flux | ZIB | Krea2]" - those are separate files for separate + architectures shipped under one model page, not a combined adapter. + • To combine two full Krea 2 checkpoints rather than adapters, ComfyUI + ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per- + section weights for first., tmlp., txtmlp., tproj., + txtfusion.projector, txtfusion.layerwise_blocks.0-1, + txtfusion.refiner_blocks.0-1, blocks.0-27 and last. + • The author never states which weight family this targets. Silent + degradation, never a hard error, between Raw and Turbo. The two share + one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" + - so every LoRA tensor shape matches on both and ComfyUI/diffuser + +─────────────────────────────────────────────────────────────────────────────── + PROS +─────────────────────────────────────────────────────────────────────────────── + + Widely adopted — 7,530 downloads. [source: CivitAI stats] + + Positively received — 337 thumbs-up with no down-votes. [source: + CivitAI stats] + + Compact at 224 MB — cheap to stack with other LoRAs. [source: file + size] + +─────────────────────────────────────────────────────────────────────────────── + CONS & CAVEATS +─────────────────────────────────────────────────────────────────────────────── + - Author references NSFW output — review before using in shared work. + [source: description] + - No trigger word — the LoRA is always on once loaded, so strength is the + only control. [source: model metadata] + - Author never states which generation mode it was trained for; test + against your own workflow before relying on it. [source: description] + +─────────────────────────────────────────────────────────────────────────────── + COMPATIBILITY WARNING +─────────────────────────────────────────────────────────────────────────────── + Silent degradation, never a hard error, between Raw and Turbo. The two + share one architecture - ai-toolkit's config is commented "The reference + 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" - so + every LoRA tensor shape matches on both and ComfyUI/diffusers will load an + adapter against either without complaint. The lvladikov 4-step LoRA card + states this for the reverse direction: "Every layer it targets also exists + in Krea 2 Raw, so it will load there without complaint - but that is a side + effect of the shared architecture, not a supported mode", and reports + mixed, subject-dependent results with malformed subjects (duplicated heads, + fused limbs) at low step counts, needing 14+ steps with Raw's CFG on before + output is coherent. The intended direction (train on Raw, run on Turbo) is + officially blessed. The more common practical failure is a settings + mismatch rather than a weight mismatch: running a Turbo-targeted LoRA at + Raw's 52 steps + CFG 3.5, or leaving CFG enabled on Turbo, degrades output + with no warning. Loading a Krea 2 LoRA against a different model family + entirely (Flux, Qwen-Image, Z-Image) is also not an error - ComfyUI logs + "lora key not loaded: {key}" per key and generates as if no LoRA were + applied. + +─────────────────────────────────────────────────────────────────────────────── + AUTHOR'S DESCRIPTION +─────────────────────────────────────────────────────────────────────────────── + Ray Versatile Face： + This design aims to enhance the facial features of Asian women, making them + more diverse. If you're tired of the typical "Instagram face" or pretty + young women, try looking at Asian women with personality and international + flair. + 用于强化亚裔女性的脸，使其更加多样性。看够了网红脸或者漂亮的年轻女性，也可以试试有个性和国际化的亚裔女性。 + Ray Surreal Style: + Surrealism, visuals, atmosphere, and aesthetics. + 超现实主义、视觉、氛围、美学。 + Ray Niji style: + Niji6 style, semi-realistic style, illustration style + niji6风格，半写实风格，插画风格 + Ray Lotus Blossoms： + yes, you read that right, 0.15-0.8. 是的，你没看错就是0.15-0.8。 + It enhances East Asian women, stabilizes limbs, and influences all styles. + The raw image results are very similar to those from ZImage. + 增强东亚女性，稳定肢体，影响所有风格。生图结果和zimage很相似。 + Ray Skin detail krea2 v1： + Used to enhance skin details, with NSFW correction capabilities. + 用于提升皮肤细节, 含nsfw修正能力。 + Pantyhose fix krea2 v1： + Krea2 had a problem confusing terms like pantyhose, stockings, thigh-high + stockings, and socks, so I trained this Lora to solve that. Now you can + describe your stockings and pantyhose in more detail, including: thin, + thick, semi-sheer, vertical stripes, horizontal stripes, glossy, matte, + shiny, up to the waist, elastic waistband, center seam, shaping, reinforced + toes, and some other terms not listed or that I don't understand. You can + try them out too. + krea2有个问题，会把连裤袜，丝袜，大腿袜，袜子，这些词搞混，所以我训练了这个lora来解决这个问题。你现在可以更详细的描述你的丝袜和连裤袜了，包括 + 了：薄的，厚的，半透明，竖条纹，横条纹，亮面，哑光，闪亮，至腰部，弹性腰带，中缝，塑性，加固脚趾，还有一些没列举或者我不懂的术语你也可以试试。 + Photography style krea2 v1： + It influences a wide range of styles without a specific specialty, serving + primarily to enhance the visual impact of the image. Starting your prompt + with "This highly visually striking, high-contrast poster," will yield + results closer to the "Ray Artshoot ZIB" style. + 影响的风格较多，无特长，仅增强画面，提示词开头以："This highly visually striking, high-contrast + poster,"，效果会更接近于Ray Artshoot ZIB。 + No trigger keywords，无触发词，weight权重0.4-1.2，recommended推荐：1 + cinematic film style krea2 v1： + Kodak Vision3 250D cinematic film style; Leica-esque aesthetic. Tailored to + Asian tastes, with East Asian enhancements and Chinese elements. + Kodak Vision3 250D 电影胶片风格，徕卡风格。符合亚洲审美，东亚强化，中国元素。 + No trigger keywords，无触发词，weight权重0.4-1.2 + Hisoul style krea2 v1： + Higgsfield Soul Style, very artistic, Its unique character is used to + enhance realism and aesthetics. + Higgsfield Soul风格，气质非常独特，用于增加真人感和增强美学。 + No trigger keywords，无触发词，weight权重0.4-1.2 + Fujifilm style krea2 v1： + Japanese film photography style, enhanced for Asian subjects, Japanese + elements. Increases realism and stylization. + 日系胶片摄影风格，亚裔强化，日本元素。增加真实感和风格化。 + No trigger keywords，无触发词，weight权重0.5-1.2 + TensorArt review process is becoming increasingly strict, Therefore, I will + gradually move my new models to civitai and seeart. Reproduction is + permitted, but please indicate the source. Thank you for your attention. + Tensorart 的审核日益严格，所以我会把我的新模型陆续搬到civitai和seeart。可以转载，转载请注明出处。感谢关注。 + For online use, please visit TensorArt | Seaart + 如需在线使用可以到 TenserArt | Seaart + +─────────────────────────────────────────────────────────────────────────────── + VERSION NOTES +─────────────────────────────────────────────────────────────────────────────── + No trigger keywords，无触发词，weight, 权重：0.5-1.2 + +─────────────────────────────────────────────────────────────────────────────── + EXAMPLE MEDIA +─────────────────────────────────────────────────────────────────────────────── + 6 preview(s) in ./example_images/ — community results for + this LoRA. Videos keep a matching _poster.jpg still frame. + + No prompts are recorded below: CivitAI's public API flags these posts as + having generation metadata but does not return it, so the prompts cannot be + scraped. Open the model page to read them. + + 01_image.jpeg (image, 1024x1536) + 02_image.jpeg (image, 1024x1536) + 03_image.jpeg (image, 1024x1536) + 04_image.jpeg (image, 1024x1536) + 05_image.jpeg (image, 1024x1536) + 06_image.jpeg (image, 1024x1536) + +─────────────────────────────────────────────────────────────────────────────── + HOW TO USE (COMFYUI) +─────────────────────────────────────────────────────────────────────────────── + 1. Update ComfyUI to at least v0.26.0 (released 2026-06-23), which added + local Krea2 support. For reference-image / style-reference / edit LoRAs you + need at least v0.29.0 (2026-07-29), which added regular and timestep-zero + reference images for Krea 2. ModelMergeKrea2 arrived in v0.27.0 + (2026-06-30). + + 2. Put the base stack in place (files from huggingface.co/Comfy-Org/Krea-2, + which is not gated): + - ComfyUI/models/diffusion_models/krea2_turbo_fp8_scaled.safetensors + (documented recommended default; the repo also holds krea2_turbo_bf16, + _int8_convrot, _mxfp8, _nvfp4 and krea2_raw_bf16 / _fp8_scaled / + _int8_convrot) + - ComfyUI/models/text_encoders/qwen3vl_4b_fp8_scaled.safetensors (bf16 + also available) + - ComfyUI/models/vae/qwen_image_vae.safetensors + + 3. Put the LoRA .safetensors in ComfyUI/models/loras/ - flat, no per-model + subfolder required. + + 4. Build or load the graph. The official image_krea2_turbo_t2i.json + template is: UNETLoader (krea2_turbo_fp8_scaled.safetensors) -> + LoraLoaderModelOnly -> KSampler; CLIPLoader + (qwen3vl_4b_fp8_scaled.safetensors, type "krea2") -> text encode -> + KSampler positive, with ConditioningZeroOut into negative; VAELoader + (qwen_image_vae.safetensors) -> VAEDecode -> SaveImage; EmptyLatentImage + 1024x1024 driven by ResolutionSelector. An int8 sibling template + (image_krea2_turbo_t2i_int8.json) also ships. + + 5. Load the LoRA with LoraLoaderModelOnly, not LoraLoader - both official + templates do. Krea 2 LoRAs patch the DiT only; note that the DiT's own + text_fusion / txtfusion blocks are a normal LoRA target (the official style + LoRAs patch them), but the Qwen3-VL text encoder itself is not. Set + strength 1.0 for Krea's official style LoRAs (the shipped template uses + 0.8) and add the LoRA's trigger word to the prompt. + + 6. Sampler settings: steps 8, cfg 1, sampler euler, scheduler simple, + denoise 1.0. Do not add a ModelSampling node for plain t2i - ComfyUI + already applies shift 1.15 from Krea2.sampling_settings. + + 7. For style reference, use the dedicated template: UNETLoader + krea2_turbo_int8_convrot.safetensors, LoraLoaderModelOnly + krea2_style_reference.safetensors at 1.0, ModelSamplingFlux (1.15, 0.5), + TextEncodeQwenImageEditPlus fed by LoadImage, + FluxKontextMultiReferenceLatentMethod set to index_timestep_zero, and + SamplerCustomAdvanced with KSamplerSelect euler + BasicScheduler simple/8 + + CFGGuider cfg 1. + + 8. For instruction editing (Krea 2 Identity Edit), stock nodes are not + enough: install the ComfyUI-Krea2Edit node pack, which ships its own + workflows, and run Turbo at 8-12 steps / cfg 1 (or Raw at 20 steps / cfg 3 + for removals) with LoRA strength 1.0. + + 9. Low-VRAM alternative: install city96's ComfyUI-GGUF and use a community + GGUF build (realrebelai/KREA-2_GGUFs, quantized from Turbo; + molbal/krea2-gguf, quantized from Raw) with the Unet Loader (GGUF) node in + place of UNETLoader. There is no official GGUF release. LoRA loading is + unchanged. + +─────────────────────────────────────────────────────────────────────────────── + Scraped by lorakit on 2026-09-05 00:16 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/ray-krea2-style-loras/metadata.json b/ray-krea2-style-loras/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..b786121e60d889312ca9da262b537d439fe34a1f --- /dev/null +++ b/ray-krea2-style-loras/metadata.json @@ -0,0 +1,133 @@ +{ + "name": "Ray Krea2 Style Loras", + "slug": "ray-krea2-style-loras", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 2802215, + "version_id": 3280920, + "version_name": "Ray Versatile Face", + "source": "civitai", + "url": "https://civitai.com/models/2802215", + "repo_id": null, + "file_path": null, + "author": "Ray3780", + "category": { + "key": "detail", + "label": "Detail / anatomy fixer" + }, + "modalities": [], + "primary_modality": "unspecified", + "checkpoint_family": "unknown", + "trigger_words": [], + "recommended_settings": {}, + "chaining": { + "author": { + "stacks": "not stated", + "notes": "", + "avoid": [], + "requires": [], + "order": "", + "strength": "", + "stated": false + }, + "general_rules": [ + "The most important rule is not about chaining but about which checkpoint you chain onto: the official README says in bold \"TRAIN on Raw and RUN on Turbo\", and \"We highly recommend using RAW for training LoRAs and applying them on Turbo for inference.\" LoRAs trained on Raw are stated to work well on Turbo.", + "Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2 LoRAs are DiT-only, so there is no CLIP side to chain and the text-encoder branch stays untouched however many LoRAs you add.", + "Krea 2 LoRAs stack in practice and the community treats multi-LoRA as routine - style + character, style + slider, realism + character. Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one writes \"I recently stacked over 20 of my own LoRAs on a single generation and the output still came out great.\"", + "Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is loaded. The CivitAI 12GB training guide also notes rank affects composability: \"rank 8 produced a lighter style influence that was easier to stack with other LoRAs, rank 16 provided a useful balance between concept strength and composability, and rank 32 was better suited to a precise object or concept intended to take priority over other stacked LoRAs.\"", + "Accelerator LoRAs are position-sensitive in intent, not in node order: a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to-Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is explicitly recommended by the 4-step LoRA's author.", + "Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card says \"Keep mu = 1.15. The training targets are anchored to that grid; a different shift moves the [sampler off them]\"; the same reasoning applies to any Turbo-trained adapter in the chain.", + "Reference/edit LoRAs are designed to be stacked under ordinary subject and style LoRAs - the Identity Edit card states \"Composes with your LoRAs: character/body/style LoRAs stack on top and steer the prior\" and recommends a stacked subject LoRA when unusual subjects drift toward the base prior.", + "Krea 2 LoRAs cannot be chained with LoRAs for any other base model. Many CivitAI pages carry multi-model titles like \"(Krea-2 + ZIT)\" or \"[Flux | ZIB | Krea2]\" - those are separate files for separate architectures shipped under one model page, not a combined adapter.", + "To combine two full Krea 2 checkpoints rather than adapters, ComfyUI ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per-section weights for first., tmlp., txtmlp., tproj., txtfusion.projector, txtfusion.layerwise_blocks.0-1, txtfusion.refiner_blocks.0-1, blocks.0-27 and last.", + "The author never states which weight family this targets. Silent degradation, never a hard error, between Raw and Turbo. The two share one architecture - ai-toolkit's config is commented \"The reference 'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)\" - so every LoRA tensor shape matches on both and ComfyUI/diffuser" + ] + }, + "pros": [ + "Widely adopted — 7,530 downloads. [source: CivitAI stats]", + "Positively received — 337 thumbs-up with no down-votes. [source: CivitAI stats]", + "Compact at 224 MB — cheap to stack with other LoRAs. [source: file size]" + ], + "cons": [ + "Author references NSFW output — review before using in shared work. [source: description]", + "No trigger word — the LoRA is always on once loaded, so strength is the only control. [source: model metadata]", + "Author never states which generation mode it was trained for; test against your own workflow before relying on it. [source: description]" + ], + "file": { + "name": "Ray Versatile Face v1.safetensors", + "size_bytes": 234702984, + "size_mb": 223.83, + "sha256": "89debcdae181477684af446c8208a39a40612ba2dd4e3e6f53808ecd093e85d2" + }, + "stats": { + "downloads": 7530, + "thumbs_up": 337, + "thumbs_down": 0 + }, + "popularity": { + "downloads": 7530, + "thumbs_up": 337, + "thumbs_down": 0, + "comments": 17, + "approval": "100% positive (337/337)", + "age_days": 4.9, + "downloads_per_day": 66.88, + "rank_by_downloads": 77, + "rank_by_momentum": 41, + "of_total": 150, + "adoption_tier": "Modest", + "momentum": "Steady · fresh release" + }, + "selection_score": 11574.0, + "selection_reasons": [ + "7,530 downloads", + "337 thumbs-up", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "1024x1536", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "1024x1536", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "1024x1536", + "prompt": "" + }, + { + "file": "04_image.jpeg", + "type": "image", + "resolution": "1024x1536", + "prompt": "" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "1024x1536", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "1024x1536", + "prompt": "" + } + ], + "tags": [ + "style", + "aesthetic", + "krea2", + "slyle" + ], + "scraped_at": "2026-09-05T00:16:12+00:00", + "path": "krea-2/ray-krea2-style-loras" +} diff --git a/yfg-chatgpt-4o-style-flux-zit-krea2/YFG-ChatGPT_4o-Krea2_3k-v1.safetensors b/yfg-chatgpt-4o-style-flux-zit-krea2/YFG-ChatGPT_4o-Krea2_3k-v1.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..faaa59736e72077d7c7046b851e08b875a867512 --- /dev/null +++ b/yfg-chatgpt-4o-style-flux-zit-krea2/YFG-ChatGPT_4o-Krea2_3k-v1.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e92fb6e599d1fea0cee186c658f67c9d55c76918891c0c95ee692edf088a2123 +size 228587808 diff --git a/yfg-chatgpt-4o-style-flux-zit-krea2/example_images/01_image.jpeg b/yfg-chatgpt-4o-style-flux-zit-krea2/example_images/01_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..bd23b3ee9d38991eebd044ddbb062c02f471e142 --- /dev/null +++ b/yfg-chatgpt-4o-style-flux-zit-krea2/example_images/01_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c21e0e69ef4acb7273c990704a1a55062cd29511001e94568e3fdac9fca5b0b +size 4501391 diff --git a/yfg-chatgpt-4o-style-flux-zit-krea2/example_images/02_image.jpeg b/yfg-chatgpt-4o-style-flux-zit-krea2/example_images/02_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..933b8081c801dc0c8f99d9b3f67b8951102adeb2 --- /dev/null +++ b/yfg-chatgpt-4o-style-flux-zit-krea2/example_images/02_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9cad5d81decd3c5056fc6b70604656df8bbcf28b6a5925553b8b02ef80422dfd +size 3704660 diff --git a/yfg-chatgpt-4o-style-flux-zit-krea2/example_images/03_image.jpeg b/yfg-chatgpt-4o-style-flux-zit-krea2/example_images/03_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..069c93fd05e8140c9e5c5cbdd7b17919ab6cbcec --- /dev/null +++ b/yfg-chatgpt-4o-style-flux-zit-krea2/example_images/03_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d2b301de5f2e507a31cfad5e624fcf7ad066e8bd9d34b5207b14647017cbc35 +size 3478040 diff --git a/yfg-chatgpt-4o-style-flux-zit-krea2/example_images/04_poster.jpg b/yfg-chatgpt-4o-style-flux-zit-krea2/example_images/04_poster.jpg new file mode 100644 index 0000000000000000000000000000000000000000..c3f77e97c2d1512a68b49ea4608056cf3ae4d870 Binary files /dev/null and b/yfg-chatgpt-4o-style-flux-zit-krea2/example_images/04_poster.jpg differ diff --git a/yfg-chatgpt-4o-style-flux-zit-krea2/example_images/04_video.mp4 b/yfg-chatgpt-4o-style-flux-zit-krea2/example_images/04_video.mp4 new file mode 100644 index 0000000000000000000000000000000000000000..28d92aea8b867eb59b75d40a159e5dcd4f2fd670 --- /dev/null +++ b/yfg-chatgpt-4o-style-flux-zit-krea2/example_images/04_video.mp4 @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd51a6b5bae6e63f0e5d45f16630b6eeff06c8f08feae9f769975d463ded6454 +size 3220470 diff --git a/yfg-chatgpt-4o-style-flux-zit-krea2/example_images/05_image.jpeg b/yfg-chatgpt-4o-style-flux-zit-krea2/example_images/05_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..ffe3ce65d4b01fb9c973a38a69cff90238edbe14 --- /dev/null +++ b/yfg-chatgpt-4o-style-flux-zit-krea2/example_images/05_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3e3251ee084a7c0a0fe42c46f2351f453bd5e88ac095398a7861e179511916ae +size 3947705 diff --git a/yfg-chatgpt-4o-style-flux-zit-krea2/example_images/06_image.jpeg b/yfg-chatgpt-4o-style-flux-zit-krea2/example_images/06_image.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..207aa50cc2371bab26f0435d41300972fa2193b4 --- /dev/null +++ b/yfg-chatgpt-4o-style-flux-zit-krea2/example_images/06_image.jpeg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:978f0c5f45b5ab3703d934a7df42b01de87ecf3bbe9aae4aad86acfa7bcbe947 +size 8847983 diff --git a/yfg-chatgpt-4o-style-flux-zit-krea2/info.txt b/yfg-chatgpt-4o-style-flux-zit-krea2/info.txt new file mode 100644 index 0000000000000000000000000000000000000000..be847b025c526ce29001ae8ff483566548cae67d --- /dev/null +++ b/yfg-chatgpt-4o-style-flux-zit-krea2/info.txt @@ -0,0 +1,344 @@ +═══════════════════════════════════════════════════════════════════════════════ + YFG CHATGPT 4O STYLE [FLUX | ZIT | KREA2] +═══════════════════════════════════════════════════════════════════════════════ + +─────────────────────────────────────────────────────────────────────────────── + IDENTITY +─────────────────────────────────────────────────────────────────────────────── + Base model Krea 2 + Generation mode Text-to-image + Checkpoint Krea 2 RAW (krea/Krea-2-Raw) + Category Visual style + Author gonzalu + Source https://civitai.com/models/1470117 + Provenance CivitAI · model 1470117 · version 3085008 + +─────────────────────────────────────────────────────────────────────────────── + WEIGHTS +─────────────────────────────────────────────────────────────────────────────── + File YFG-ChatGPT_4o-Krea2_3k-v1.safetensors + Version Krea2 3K - v1.0 + Size 218.0 MB + Format SafeTensor / unknown precision + SHA256 e92fb6e599d1fea0cee186c658f67c9d55c76918891c0c95ee692edf088a2123 + Published 2026-06-30 + +─────────────────────────────────────────────────────────────────────────────── + TRIGGER WORDS +─────────────────────────────────────────────────────────────────────────────── + › "YFG-4o style" + +─────────────────────────────────────────────────────────────────────────────── + 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 5,768 + Rating 100% positive (485/487) + Comments 0 + Published 66 days ago + Download rate 9.5/day since release + + Adoption rank #100 of 150 in this collection (Modest) + Momentum rank #142 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 +─────────────────────────────────────────────────────────────────────────────── + + Widely adopted — 5,768 downloads. [source: CivitAI stats] + + Positively received — 485 thumbs-up, 2 down. [source: CivitAI stats] + + Explicit trigger word(s) — YFG-4o style — so the effect can be turned + on and off from the prompt. [source: model metadata] + + Compact at 218 MB — cheap to stack with other LoRAs. [source: file + size] + +─────────────────────────────────────────────────────────────────────────────── + 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 +─────────────────────────────────────────────────────────────────────────────── + YFG ChatGPT-4o Style: Dark Cinematic Mood, Imagery and Tasteful Grain YFG + ChatGPT-4o Style is a LoRA that channels the moody, filmic atmosphere + characteristic of the new ChatGPT-4o text-to-image aesthetic. With subdued + lighting, subtle grain, and a cinematic flair, it’s perfect for injecting + atmospheric depth and drama into your scenes. + Key Features - Dark, Cinematic Tones + - Emphasizes low-key lighting, strong shadows, and a dramatic contrast that + evokes film noir or gritty movie stills. + + - Great for narrative-driven or emotionally intense imagery. + + - Subtle, Tasteful Grain + - Adds a refined grain texture reminiscent of high-quality film stock. + + - Contributes an understated retro or documentary vibe without overwhelming + details. + + - Flexible Strength + - 0.3 – 1.2 tested, with 0.80 as the sweet spot for balanced cinematic + flair. + + - Lower for mild atmospheric elements, higher for a more striking filmic + transformation. + + - No Specific Trigger Word + - Simply describe the mood and style—e.g., “in the style of YFG + ChatGPT-4o”—to bring out this aesthetic. + + - Works well with subject-focused prompts, thematic keywords like “moody,” + “low-key lighting,” or “intense shadows.” + + Recommended Usage - LoRA Strength + - 0.3 – 0.6 : Subtle, film-like grading with moderate shadows. + + - 0.8 : Balanced cinematic mood, refined texture, and distinct dark tones. + + - 1.0+ : Maximum transformation—heavy shadows, pronounced grain, and deeply + atmospheric. + + - Sample Prompt Ideas + - “A dramatic portrait in a dimly lit room, capturing moody expressions and + subtle grain in a YFG ChatGPT-4o style.” + + - “A nighttime cityscape with neon reflections on wet pavement, cinematic + lighting, and refined film grain.” + + Tips & Tricks - Embrace Low-Key Lighting + - Mention “soft overhead lighting,” “dim corridor,” “noir lighting,” or + “dramatic backlight” to accentuate shadows and highlight texture. + + - Color Palettes + - While dark scenes are typical, you can guide color usage for + variety—“cool, bluish shadows,” “warm tungsten glow,” or “monochrome filmic + aesthetic.” + + - Combine With Other Styles + - Layer with a subject LoRA (e.g., character design, architecture, etc.) + for moody backdrops or narrative context. + + - Tweak each model’s strength to prevent overshadowing key details. + + - Refining the Grain + - If the grain is too heavy, lower LoRA strength or prompt for “slight/no + visible grain.” + + - If you want more texture, incorporate prompts like “noticeable film + grain,” “retro camera effect,” or “gritty cinematic vibe.” + + YFG ChatGPT-4o Style captures the brooding ambiance of modern cinema, + combining deep contrasts, filmic grain, and low-key lighting to deliver + images brimming with drama and atmosphere. Experiment with the LoRA’s + strength to find just the right balance of gloom and glam for your creative + vision! + +─────────────────────────────────────────────────────────────────────────────── + 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, 1248x1872) + 02_image.jpeg (image, 1248x1872) + 03_image.jpeg (image, 1248x1872) + 04_video.mp4 (video, 416x720) + still: 04_poster.jpg + 05_image.jpeg (image, 1248x1872) + 06_image.jpeg (image, 2000x3000) + +─────────────────────────────────────────────────────────────────────────────── + 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 00:16 UTC. + Stats and description are the author's; pros/cons are derived from the + evidence tagged beside each line. +─────────────────────────────────────────────────────────────────────────────── diff --git a/yfg-chatgpt-4o-style-flux-zit-krea2/metadata.json b/yfg-chatgpt-4o-style-flux-zit-krea2/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..dda3ba8171bc6fcada36e9b7eb519084050db9fa --- /dev/null +++ b/yfg-chatgpt-4o-style-flux-zit-krea2/metadata.json @@ -0,0 +1,139 @@ +{ + "name": "YFG ChatGPT 4o Style [Flux | ZIT | Krea2]", + "slug": "yfg-chatgpt-4o-style-flux-zit-krea2", + "base_model": "Krea 2", + "media_type": "image", + "model_id": 1470117, + "version_id": 3085008, + "version_name": "Krea2 3K - v1.0", + "source": "civitai", + "url": "https://civitai.com/models/1470117", + "repo_id": null, + "file_path": null, + "author": "gonzalu", + "category": { + "key": "style", + "label": "Visual style" + }, + "modalities": [ + "t2i" + ], + "primary_modality": "t2i", + "checkpoint_family": "raw", + "trigger_words": [ + "YFG-4o style" + ], + "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.", + "Undistilled base checkpoint, no step or CFG distillation and no post-training. The card says \"its not recommended for inference use. Its a good base for finetuning or post-training\". Official settings: 52 steps, CFG 3.5, trained up to 1K. This is where Krea tells you to train LoRAs." + ] + }, + "pros": [ + "Widely adopted — 5,768 downloads. [source: CivitAI stats]", + "Positively received — 485 thumbs-up, 2 down. [source: CivitAI stats]", + "Explicit trigger word(s) — YFG-4o style — so the effect can be turned on and off from the prompt. [source: model metadata]", + "Compact at 218 MB — cheap to stack with other LoRAs. [source: file size]" + ], + "cons": [], + "file": { + "name": "YFG-ChatGPT_4o-Krea2_3k-v1.safetensors", + "size_bytes": 228587808, + "size_mb": 218.0, + "sha256": "e92fb6e599d1fea0cee186c658f67c9d55c76918891c0c95ee692edf088a2123" + }, + "stats": { + "downloads": 5768, + "thumbs_up": 485, + "thumbs_down": 2 + }, + "popularity": { + "downloads": 5768, + "thumbs_up": 485, + "thumbs_down": 2, + "comments": 0, + "approval": "100% positive (485/487)", + "age_days": 66.4, + "downloads_per_day": 9.52, + "rank_by_downloads": 100, + "rank_by_momentum": 142, + "of_total": 150, + "adoption_tier": "Modest", + "momentum": "Long tail" + }, + "selection_score": 11988.0, + "selection_reasons": [ + "5,768 downloads", + "485 thumbs-up", + "documented triggers + usage notes", + "top-ranked for its category" + ], + "example_media": [ + { + "file": "01_image.jpeg", + "type": "image", + "resolution": "1248x1872", + "prompt": "" + }, + { + "file": "02_image.jpeg", + "type": "image", + "resolution": "1248x1872", + "prompt": "" + }, + { + "file": "03_image.jpeg", + "type": "image", + "resolution": "1248x1872", + "prompt": "" + }, + { + "file": "04_video.mp4", + "type": "video", + "resolution": "416x720", + "prompt": "", + "poster": "04_poster.jpg" + }, + { + "file": "05_image.jpeg", + "type": "image", + "resolution": "1248x1872", + "prompt": "" + }, + { + "file": "06_image.jpeg", + "type": "image", + "resolution": "2000x3000", + "prompt": "" + } + ], + "tags": [ + "cinematic", + "moody", + "dark tones", + "tasteful grain", + "chatgpt-4o", + "style", + "chatgpt" + ], + "scraped_at": "2026-09-05T00:16:00+00:00", + "path": "krea-2/yfg-chatgpt-4o-style-flux-zit-krea2" +}