scraed/LanPaint

What it solves

LanPaint provides a universal, training-free inpainting and outpainting solution for diffusion models. It eliminates the need for model-specific inpainting training or complex blending workarounds, allowing users to generate high-quality, entirely new content within masked areas across a wide variety of image and video models.

How it works

LanPaint introduces a specialized sampler that implements a "Think Mode." This allows the model to perform multiple iterations of internal processing (thinking) before the final denoising process begins. By investing more computation time through these iterations, the sampler achieves asymptotically exact and fast conditional sampling, resulting in superior inpainting quality without requiring any changes to the model's weights.

Who it’s for

It is primarily designed for digital artists and AI researchers using ComfyUI who want high-fidelity inpainting and outpainting across different models (such as Flux, SDXL, Wan 2.2, and MiniMax H3) without having to train custom inpainting versions of those models.

Highlights

• Universal Compatibility: Works with a vast array of models including Ideogram4, Krea2, Z-image, Hunyuan, Wan 2.2, Qwen, Anima, SD 3.5, Flux, and SDXL.
• Training-Free: Operates out-of-the-box with existing models without requiring additional training or fine-tuning.
• Multimodal Inpainting: Supports simultaneous video and audio inpainting using the MiniMax H3 pipeline.
• Integrated Video Mask Editor: Includes a built-in editor for painting per-frame masks with automatic interpolation between keyframes.
• Temporal Consistency: Maintains consistency across video frames for tasks like video inpainting and outpainting.