scraed/LanPaint

What it solves

LanPaint provides a universal, training-free way to perform high-quality image and video inpainting and outpainting across a wide variety of pretrained diffusion and rectified-flow models. It eliminates the need for model fine-tuning or backpropagation, allowing users to perform mask-constrained local editing, object replacement, and generative fill on almost any existing model.

How it works

LanPaint implements a partial conditional sampler that allows the model to "think" through multiple iterations before denoising. This "Think Mode" enables the investment of more computation time to improve inpainting quality. It works as a a ComfyUI extension and supports a variety of modalities, including images (with transparency support for Qwen 2.1) and video/audio (via a nested latent pipeline for MiniMax H3).

Who it’s for

It is designed for digital artists and AI researchers using ComfyUI who want to perform precise local edits on images or videos without having to train specific inpainting models for every base model they use.

Highlights

• Universal Compatibility: Works with a vast array of models including Flux, SDXL, SD 1.5, Wan 2.2, Qwen Image, and MiniMax H3.
• Training-Free: No fine-tuning or backpropagation required; works out of the box with pretrained models.
• Multi-modal Inpainting: Supports image inpainting, outpainting, video inpainting, and even combined video and audio inpainting.
• Integrated Video Mask Editor: Includes a built-in editor for painting per-frame masks with automatic interpolation and the ability to export masks into video metadata.
• Think Mode: Allows for iterative refinement to achieve superior inpainting quality by increasing computation time.