177 lines
6.6 KiB
JSON
177 lines
6.6 KiB
JSON
{
|
|
"display_name": "Krea 2 - RAW",
|
|
"model_name": "Krea 2 RAW",
|
|
"version": "RAW",
|
|
"architecture": "krea2",
|
|
"family": "krea2_raw",
|
|
"author": "Krea AI",
|
|
"description": "Krea 2 RAW — the undistilled 12B base checkpoint (full-step, real cfg). Krea ships it for fine-tuning/LoRA training, but because it is NOT step-distilled it re-details cleanly at a LOW denoise — which makes it the right model for a faithful, blotch-free TILED upscale (the distilled Turbo can't: low denoise blotches, high denoise seams). Same Qwen3-VL encoder + Qwen Image VAE as Turbo.",
|
|
"license": "Krea 2 Community License Agreement v.1 (open weights; commercial use requires an Enterprise License once total annual revenue reaches $1M)",
|
|
"tags": [
|
|
"krea",
|
|
"krea2",
|
|
"dit",
|
|
"transformer",
|
|
"photorealistic",
|
|
"raw",
|
|
"base",
|
|
"undistilled"
|
|
],
|
|
"url": "https://huggingface.co/Comfy-Org/Krea-2",
|
|
"release_date": "2026-06-18",
|
|
"tag_sources": [],
|
|
"tag_format": "spaces",
|
|
"prompt_style": "natural",
|
|
"trigger": "",
|
|
"default_prompt_id": "krea2-default",
|
|
"negative": "",
|
|
"notes": "Undistilled base: full-step sampling with a REAL cfg. Render recipe (52 steps / cfg 3.5 / er_sde+simple) is DOCUMENTED — Krea's official CLI is `--checkpoint oss_raw --steps 52 --cfg 3.5`. Negatives use the real CFG branch but efficacy on the Qwen3-VL encoder is A/B-pending (may be inert like Qwen-Image). Shift is baked into the model (no ModelSamplingAuraFlow needed; mu~1.15 at 1024). Shares qwen3vl_4b + qwen_image_vae with Turbo. ~13 GB fp8 / ~26 GB bf16. Slow (no distillation) — for UPSCALE re-detail the JS engine uses fewer steps at low denoise, this 52-step recipe is for from-scratch t2i. RESOLUTION: RAW is DOCUMENTED 'up to 1024' — keep BOTH dims <=1024 (1024x1024 is its sweet spot; long edge >1024 like 1152/1216 melts faces — confirmed). Turbo is the 1k-2k model. For bigger output, render <=1024 then upscale (RAW tiles at 1024 = exactly its trained size, why the tiled upscale is so clean). SHIFT FIX (critical): ComfyUI hardcodes flow-shift mu=1.15 (Turbo's value) for the WHOLE krea2 class, but RAW needs RESOLUTION-DYNAMIC shifting (mu~0.906 @ 1024^2, smaller below) — the over-shift compresses the detail tail and melts faces. The raw t2i/i2i templates insert a ModelSamplingFlux node (base_shift 0.5, max_shift 0.90625, width/height = render size) which REPLACES the bad baked shift (verified against ComfyUI master: supported_models.py Krea2 shift=1.15, ModelType.FLUX). SHIP RECIPE (quality-research-tuned): 52 steps / cfg 5.0 / er_sde / beta — DOCUMENTED crispest-photoreal combo; euler/simple is doubly-soft (euler smooths micro-texture, simple under-weights the detail band). cfg 5.0 ~ Krea --cfg 3.5 in ComfyUI's (1+g) convention + diffusers' native 4.5. ModelSamplingFlux width/height MUST match render size or mu over-shifts on non-square buckets (1024^2->0.906, 896x1024->0.852, 832x1024->0.825, 768x1024->0.798). Turbo needs no shift node (its real mu IS 1.15). Without the ModelSamplingFlux node a RAW render melts.",
|
|
"files": [
|
|
{
|
|
"label": "Diffusion Model",
|
|
"folder": "diffusion_models",
|
|
"variants": [
|
|
{
|
|
"precision": "fp8",
|
|
"filename": "krea2_raw_fp8_scaled.safetensors",
|
|
"size_bytes": 13141730784,
|
|
"source": {
|
|
"provider": "huggingface",
|
|
"repo": "Comfy-Org/Krea-2",
|
|
"path": "diffusion_models/krea2_raw_fp8_scaled.safetensors"
|
|
},
|
|
"quick_hash": null,
|
|
"sha256": null
|
|
},
|
|
{
|
|
"precision": "bf16",
|
|
"filename": "krea2_raw_bf16.safetensors",
|
|
"size_bytes": 26283332608,
|
|
"source": {
|
|
"provider": "huggingface",
|
|
"repo": "Comfy-Org/Krea-2",
|
|
"path": "diffusion_models/krea2_raw_bf16.safetensors"
|
|
},
|
|
"quick_hash": null,
|
|
"sha256": null
|
|
}
|
|
]
|
|
},
|
|
{
|
|
"label": "Text Encoder",
|
|
"folder": "text_encoders",
|
|
"variants": [
|
|
{
|
|
"precision": "fp8",
|
|
"filename": "qwen3vl_4b_fp8_scaled.safetensors",
|
|
"size_bytes": 5242467968,
|
|
"source": {
|
|
"provider": "huggingface",
|
|
"repo": "Comfy-Org/Krea-2",
|
|
"path": "text_encoders/qwen3vl_4b_fp8_scaled.safetensors"
|
|
},
|
|
"sha256": null
|
|
},
|
|
{
|
|
"precision": "bf16",
|
|
"filename": "qwen3vl_4b_bf16.safetensors",
|
|
"size_bytes": 8875719384,
|
|
"source": {
|
|
"provider": "huggingface",
|
|
"repo": "Comfy-Org/Krea-2",
|
|
"path": "text_encoders/qwen3vl_4b_bf16.safetensors"
|
|
},
|
|
"sha256": null
|
|
}
|
|
]
|
|
},
|
|
{
|
|
"label": "VAE",
|
|
"folder": "vae",
|
|
"filename": "qwen_image_vae.safetensors",
|
|
"size_bytes": 253806246,
|
|
"source": {
|
|
"provider": "huggingface",
|
|
"repo": "Comfy-Org/Krea-2",
|
|
"path": "vae/qwen_image_vae.safetensors"
|
|
},
|
|
"sha256": null
|
|
},
|
|
{
|
|
"label": "Filter-Bypass Projector-Scale LoRA",
|
|
"folder": "loras",
|
|
"filename": "krea2_filter_bypass_projector_scale.safetensors",
|
|
"size_bytes": 268,
|
|
"optional": true,
|
|
"source": {
|
|
"provider": "huggingface",
|
|
"repo": "Beinsezii/Krea-2-Raw-Projector-Scale-LoRA-Diffusers",
|
|
"path": "pytorch_lora_weights.safetensors"
|
|
},
|
|
"sha256": null
|
|
}
|
|
],
|
|
"nodes": {
|
|
"KSampler": {
|
|
"steps": 52,
|
|
"cfg": 5.0,
|
|
"sampler_name": "er_sde",
|
|
"scheduler": "beta",
|
|
"steps_range": [
|
|
40,
|
|
60
|
|
],
|
|
"cfg_range": [
|
|
3.5,
|
|
6.0
|
|
],
|
|
"sampler_name_options": [
|
|
"er_sde",
|
|
"dpmpp_2m",
|
|
"euler",
|
|
"dpmpp_sde",
|
|
"euler_ancestral"
|
|
],
|
|
"scheduler_options": [
|
|
"beta",
|
|
"simple",
|
|
"sgm_uniform",
|
|
"normal"
|
|
]
|
|
},
|
|
"EmptySD3LatentImage": {
|
|
"width": 1024,
|
|
"height": 1024,
|
|
"_resolutions": [
|
|
{
|
|
"width": 1024,
|
|
"height": 1024
|
|
},
|
|
{
|
|
"width": 832,
|
|
"height": 1024
|
|
},
|
|
{
|
|
"width": 1024,
|
|
"height": 832
|
|
},
|
|
{
|
|
"width": 768,
|
|
"height": 1024
|
|
},
|
|
{
|
|
"width": 1024,
|
|
"height": 768
|
|
},
|
|
{
|
|
"width": 896,
|
|
"height": 1024
|
|
},
|
|
{
|
|
"width": 1024,
|
|
"height": 896
|
|
}
|
|
]
|
|
}
|
|
}
|
|
} |