moss-anypaint

MOSS AnyPaint is a functional LoRA (rank 32) created by the community for MOSS-Image-Base: white areas in the mask are regenerated, while black areas are preserved. The same input setup supports both local inpainting and expanding the image to a larger canvas; masks can be thin, curved, cross the original image boundaries, or consist of multiple disconnected regions. It is recommended to stack this with the Base 8-step DMD2 distillation LoRA, generating images in 8 steps with CFG 1, using qfloat8 quantization for the frozen DiT. The author emphasizes that standard LoRA loaders and auto-generated Diffusers code from Hugging Face cannot be used; instead, use the scripts provided in the repository or the author's ComfyUI nodes.

MOSS AnyPaint is a functional LoRA (rank 32) created by the community for MOSS-Image-Base: white areas in the mask are regenerated, while black areas are preserved. The same input setup supports both local inpainting and expanding the image to a larger canvas; masks can be thin, curved, cross the original image boundaries, or consist of multiple disconnected regions. It is recommended to stack this with the Base 8-step DMD2 distillation LoRA, generating images in 8 steps with CFG 1, using qfloat8 quantization for the frozen DiT. The author emphasizes that standard LoRA loaders and auto-generated Diffusers code from Hugging Face cannot be used; instead, use the scripts provided in the repository or the author's ComfyUI nodes.

Why it matters A single LoRA handles both inpainting within the image and outpainting to expand beyond it; masks can be of any shape and consist of multiple disconnected regions. The author also provides ComfyUI nodes.

Hugging Face likes 9 · Hugging Face trending score 9 · Hugging Face downloads 296

License: Commercial use allowed · ComfyUI: Mentioned in model card · VRAM: Not specified

Reasons

  • The LoRA and inference code are licensed under Apache-2.0; each example includes the original image, mask, result, and prompt.
  • The author has additionally released ComfyUI nodes and step-by-step workflow instructions.

Usage

  • Run hf download yijunwang2/moss-anypaint, then after installing dependencies with pip install -r requirements.txt, use example.py specifying --source, --mask, --prompt, and canvas dimensions; for outpainting, add --bbox to specify the position of the original image within the new canvas.
  • ComfyUI: Install the nodes according to the instructions at github.com/alexw5702-afk/moss-anypaint.

Risks

  • The ComfyUI nodes come from a third-party GitHub repository and have not been audited by this column; please review the code yourself before installation.
  • Requires a CUDA GPU that supports BF16; when generating large areas, unrelated objects may appear, and unmasked regions may also undergo subtle changes after passing through the VAE.

Open

抱脸 image-to-image 上升榜

Explore the latest AI Radar updates