Repaint one small region of a picture, a prop or an accessory, and leave everything else alone. About 27 seconds a pass on our card.

Required models

• Diffusion model: RedCraft 3.0 Krea 2, the INT8 convrot build, redcraft23_30Krea2_int8_convrot.safetensors, in models/diffusion_models. Plain Krea 2 and Krea 2 Turbo load too; RedCraft is what we tested.
• Text encoder: the Qwen3-VL 4B text encoder from the Krea 2 downloads (see the Krea 2 guide on comfyui-wiki.com); we run a community int8 ConvRot build of it (qwen3vl_4b_int8_convrot), and the bf16 or fp8 file works the same. It goes in models/text_encoders, loaded with CLIPLoader type krea2.
• VAE: the Qwen-Image VAE, qwen_image_vae.safetensors, in models/vae, from Comfy-Org on Hugging Face.

Settings

• Steps 8. It is a distilled model, more steps do not help.
• CFG 1.0, sampler er_sde, scheduler simple, denoise 0.7
• Size: 1 megapixel buckets, 896x1152, 1152x896 or 1024x1024
• Negative: a ConditioningZeroOut. At CFG 1.0 a negative prompt does nothing, so the graph does not pretend.
• Mask: read from the red channel, white means repaint

How to use it

• Crop a square around the region, about three times its size, and scale it to 1024x1024. Load it as the crop.
• Make a black and white mask of the same size, white where the new thing goes.
• Describe the crop, as a close view of that region, not the whole picture.
• To change a shape, first paint a flat blob of the new object's colour over the masked area, then run at 0.7. To change only colour or material on the same shape, skip the blob.
• Denoise 1.0 ignores the mask edge and draws its own object at its own size. Use it only to restyle in place.
• Scale the result back and paste it over the original with a feathered edge, so nothing outside the region ever goes through the VAE. Crop and paste-back happen in your image editor; the graph does the repaint.

Tips we learned the hard way

• Density beats brevity. Write a full paragraph: who or what, where, the light, the camera. A short prompt gets you a short picture.
• Say what you want, never what you do not. No negative prompt, and no SDXL habits: weight brackets and the old separator keyword do nothing here. Write "alone in the frame, no other people" instead of a negative list.
• Name the medium. "Drawn in clean cel-shaded anime line art, drawn rather than photographed" holds a drawing; "semi-realistic" alone drifts to photo.
• Stay at 1 MP. The buckets above are what the model was trained around.

Tested on ROCm (a 7900 XTX). Core ComfyUI nodes only, so CUDA builds should run it as is.

If this worked for you, a thumbs-up here helps the next person find it.