model-card.md

---
language:
- en
tags:
- comfyui
- inpainting
- stable-diffusion
- image-generation
- computer-vision
- diffusion
- image-editing
license: gpl-3.0
library_name: comfyui-extension
pipeline_tag: image-to-image
datasets:
- stable-diffusion
- stable-diffusion-xl
- stable-diffusion-3
- qwen-image
- hidream
- flux
metrics:
- image-quality
- inpainting-accuracy
- consistency
---

# LanPaint: Universal Inpainting Sampler

## Model Description

LanPaint is a universal inpainting sampler that works with any Stable Diffusion model without requiring specialized training. It introduces a "think mode" that allows models to process multiple iterations before denoising, resulting in superior inpainting quality.

- **Developed by:** LanPaint Team
- **Model type:** ComfyUI Extension (Sampler)
- **License:** GNU General Public License v3.0
- **Repository:** [GitHub](https://github.com/scraed/LanPaint)
- **Paper:** [LanPaint: Training-Free Diffusion Inpainting with Exact and Fast Conditional Inference](https://arxiv.org/abs/2502.03491)

## Intended Use

LanPaint is designed for:
- **Image Inpainting:** Fill in masked areas of images with contextually appropriate content
- **Image Outpainting:** Extend images beyond their original boundaries
- **Character Consistency:** Generate consistent character views and variations
- **Universal Compatibility:** Work with any Stable Diffusion model (SD 1.5, XL, 3.5, Flux, HiDream, Qwen-Image)

## How to Use

### Installation
1. Install ComfyUI (version > 0.3.11)
2. Install ComfyUI-Manager
3. Install LanPaint via ComfyUI-Manager or clone to `custom_nodes` folder
4. Restart ComfyUI

### Basic Usage
- Use LanPaint nodes in the "sampling" category
- Same workflow as standard ComfyUI KSampler
- Supports any mask shape, size, or position
- Configurable "thinking" steps for quality vs. speed trade-off

### Example Workflows
- Qwen Image inpainting/outpainting
- HiDream character consistency
- SD 3.5 high-quality inpainting
- Flux model inpainting (with low guidance)

## Limitations

- **Distillation Models:** Reduced performance on models like Flux.dev (use low guidance 1.0-2.0)
- **Computation Time:** More thinking steps require more processing time
- **Memory Usage:** May require significant GPU memory for high-resolution images

## Training and Technical Details

- **Training-free:** Works out-of-the-box with existing models
- **No fine-tuning required:** Universal compatibility across model architectures
- **Exact conditional inference:** Mathematically sound approach to inpainting
- **Fast implementation:** Optimized for real-world usage

## Evaluation

LanPaint has been benchmarked against traditional inpainting methods:
- VAE Encode for Inpainting
- Set Latent Noise Mask
- Standard KSampler approaches

Results show significant improvements in:
- Content consistency
- Mask boundary handling
- Overall image quality
- Character preservation

## Citation

```bibtex
@article{lanpaint2024,
  title={LanPaint: Training-Free Diffusion Inpainting with Exact and Fast Conditional Inference},
  author={LanPaint Team},
  journal={arXiv preprint arXiv:2502.03491},
  year={2024}
}
```

## Additional Resources

- [ComfyUI Documentation](https://docs.comfy.org/)
- [LanPaintBench Repository](https://github.com/scraed/LanPaintBench)
- [Blog Post](https://scraed.github.io/scraedBlog/)
- [Community Examples](https://github.com/scraed/LanPaint/tree/master/examples)

## License

This project is licensed under the GNU General Public License v3.0 - see the [LICENSE](LICENSE) file for details.