LanPaint offers a training-free diffusion inpainting solution for Stable Diffusion and Flux models, enabling precise and seamless image edits without requiring model fine-tuning. It's designed for users of ComfyUI seeking advanced inpainting capabilities, allowing for complete pixel replacement and character consistency generation.

How It Works

LanPaint introduces a novel "thinking mode" where the model iteratively aligns masked and unmasked image regions. This bidirectional evaluation ensures new content harmonizes with existing elements, leading to more coherent and accurate results than traditional methods that often "paint over" existing pixels.

Quick Start & Requirements

• Install: Via ComfyUI-Manager (search "LanPaint") or by cloning https://github.com/scraed/LanPaint.git into ComfyUI/custom_nodes.
• Prerequisites: Latest ComfyUI (>= 0.3.11), ComfyUI-Manager. For HiDream examples: specific GGUF models (clip_g_hidream, clip_l_hidream, T5 GGUF, Llama 3.1), Flux VAE.
• Setup: Requires ComfyUI installation and node integration.
• Docs: https://github.com/scraed/LanPaint

Highlighted Details

• Zero-training inpainting compatible with any SD model, ControlNet, and Flux models.
• Enables 100% new content creation in masked areas, avoiding "painting over."
• Supports character consistency generation through inpainting.
• Includes advanced parameters for fine-tuning alignment, thinking depth, and noise levels.

Maintenance & Community

• Active development with recent updates for HiDream and Flux support.
• Community contributions are encouraged via bug reports and parameter sharing.
• arXiv Paper for technical details.

Licensing & Compatibility

• No explicit license mentioned in the README.
• Compatibility with commercial use or closed-source linking is not specified.

Limitations & Caveats

LanPaint requires binary masks (0 or 1 values) and may need parameter tuning for optimal results across different models and inpainting tasks. Compatibility issues with Flux Guidance and SD 3.5 are noted as areas for future improvement.