krea2-re-75

The Realism Engine (Krea 2 v3.1, civitai model 2688234 by RazzzHF) photorealism adapter merged into comfy-org/krea-2 at strength 0.75. All credit for the adapter goes to its author; this repo only bakes it into the base weights.

The adapter is a full-form LyCORIS LoKr (kron(w1, w2), trained with ai-toolkit); the merge adds 0.75 * kron(w1, w2) to each of the 256 targeted linear weights (all 32 transformer blocks + the text-fusion blocks; the stored alpha is inert for full-form LoKr, matching ComfyUI's LoKrAdapter).

Files

File	Base	Precision	Use
krea2-raw-re0.75-bf16.safetensors	raw	bf16	training base (full precision)
krea2-raw-re0.75-fp8.safetensors	raw	fp8_e4m3fn plain	training base (fits 32 GB cards)
krea2-turbo-re0.75-fp8-scaled.safetensors	turbo	fp8_e4m3fn scaled	inference (8-step turbo)

• The plain fp8 training file casts only the transformer-block linears to fp8; first/tmlp/tproj/txtfusion/txtmlp/last and norms stay at source precision (diffusion-pipe diffusion_model_dtype='float8' layout, no scales).
• The scaled fp8 inference file uses per-tensor scales (scale = max|W| / 448, dequant fp8 * scale) with the same _quantization_metadata layout as the stock krea-2 fp8 checkpoints, so it loads natively in ComfyUI.

ComfyUI usage

Load with UNETLoader (no LoRA node — the adapter is baked into the weights):

• inference: krea2-turbo-re0.75-fp8-scaled.safetensors, CLIPLoader type krea2, VAE qwen_image_vae. Turbo: ~8 steps, cfg 1.