LanPaint: A Training-Free High-Quality Image Inpainting Tool for Stable Diffusion Models 192

What is LanPaint?

LanPaint is a high-quality image inpainting tool that requires no additional training, compatible with all Stable Diffusion-based models. It optimizes the denoising steps through a multi-iteration "thinking" process, achieving seamless and precise inpainting results.

LanPaint Features

• No Training Required: LanPaint can be used immediately with any Stable Diffusion model, including user-customized models.

• Easy Integration: Uses the same workflow as the standard ComfyUI KSampler, making it easy to get started.

• High-Quality Inpainting: Through precise conditional inference and a fast-converging Langevin dynamics framework, LanPaint achieves high-fidelity results in a small number of iterations.

How to Use LanPaint

1. Install ComfyUI: Follow the official guide to install ComfyUI.

2. Install ComfyUI-Manager: Used for convenient extension management.

3. Install the LanPaint Node:

• Search for and install LanPaint via ComfyUI-Manager.

• Or manually clone the LanPaint repository into the ComfyUI/custom_nodes folder.

• Restart ComfyUI: After loading the LanPaint node, you can find the LanPaint node under the "Sampling" category in ComfyUI.

4. Parameter Adjustment

• LanPaint_NumSteps: Indicates the number of iterations in each denoising step. For complex tasks, it is recommended to set it to 5-10.

• LanPaint_Lambda: Controls the strength of content alignment, with a recommended value of 6.0.

• LanPaint_StepSize: The step size for each iteration, with a recommended value of 0.1. For difficult tasks, it can be set to 0.5.

• LanPaint_cfg_BIG: The CFG scale used to align masked and unmasked regions, with a recommended value of 8.

LanPaint Usage

LanPaint provides a variety of examples, including restoring a basket into a basketball, a white shirt into a blue shirt, a smile into a sad expression, and more. Users can download the relevant workflows and mask files, then load the corresponding models and images in ComfyUI to try out these examples.

LanPaint Advantages

• Fast Convergence: Through the Fast Langevin Dynamics (FLD) framework, LanPaint achieves high-quality inpainting in very few iterations.

• Avoids Local Optima Traps: Through the Bi-directional Guidance (BiG) score function, LanPaint avoids local optima issues caused by large-step diffusion steps.

• Broad Applicability: Can be directly applied to various community-trained diffusion models without training on specific datasets or models.

LanPaint Use Cases

• Image Restoration: Repair old photos, remove scratches or damaged parts.

• Creative Design: Extend image boundaries, replace content, generate new elements.

• Artistic Creation: Generate stylistically consistent new content based on text prompts.

• Commercial Image Processing: Ad design, product image optimization.

• Privacy Protection: Remove mosaics, restore blurred content.

GitHub Repository:https://github.com/scraed/LanPaint