krea2-snofs-75

The snofs style adapter merged into comfy-org/krea-2 at strength 0.75.

The adapter is a full-form LyCORIS LoKr (kron(w1, w2)); the merge adds 0.75 * kron(w1, w2) to each of the 256 targeted linear weights (the stored alpha is inert for full-form LoKr, matching ComfyUI's LoKrAdapter).

Files

File	Base	Precision	Use
krea2-raw-snofs0.75-bf16.safetensors	raw	bf16	training base (full precision)
krea2-turbo-snofs0.75-fp8-scaled.safetensors	turbo	fp8_e4m3fn scaled	inference (8-step turbo)

The fp8 file uses per-tensor scaled fp8 (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 — snofs is baked into the weights):

• inference: krea2-turbo-snofs0.75-fp8-scaled.safetensors, CLIPLoader type krea2, VAE qwen_image_vae. Turbo: ~8 steps, cfg 1, sampler er_sde, scheduler simple.