⚠ Beta / test version: still being tuned, and there are some issues.

Settings

- 8 steps, CFG 1 (keep CFG at 1; negative prompts have no effect)

- euler_ancestral + beta

- 1664×2432

- load the model as fp8_e4m3fn (12 GB VRAM)

Also needed: qwen3vl_4b_fp8_scaled text encoder and qwen_image_vae.

ComfyUI workflows are included.

Prompting: long natural-language descriptions work best, not tag lists.

ComfyUI workflows: (Optional Files)

Drag a JSON onto the ComfyUI page, edit the positive prompt and press Run.

What they need

- ComfyUI: a recent version with Krea 2 support (the CLIP loader must offer type "krea2").

- Custom nodes: none. Only built-in nodes are used.

- Model files:

- models/diffusion_models/krea2_M16_furry.safetensors (this model; loaded as fp8_e4m3fn)

- models/text_encoders/qwen3vl_4b_fp8_scaled.safetensors (Krea 2 text encoder)

- models/vae/qwen_image_vae.safetensors (Krea 2 VAE)

- models/loras/krea2-masterpieces-v51.safetensors (LoRA; required by both workflows)

- GPU: about 12 GB of VRAM with fp8_e4m3fn.

Settings

- base: masterpieces LoRA at 1.5, for richer detail and contrast.

- mp10: masterpieces LoRA at 1.0, for a calmer, softer look.

- Both: 8 steps, CFG 1, euler_ancestral + beta, 1664×2432.

- Example seed: 510705642 (a wolf on a sunset balcony).

Keep CFG at 1 and steps at 8; this is a Turbo-type model, and the negative prompt has no effect.