Lanpaint: Training-Free Diffusion Inpainting with Exact and Fast Conditional Inference

• DOI: 10.48550/arxiv.2502.03491
• Publication Date: 2025-02-04
• Authors: Candi Zheng; Lan Yuan; Yang Wang
• Keywords: Inpainting

Diffusion models generate high-quality images but often lack efficient and universally applicable inpainting capabilities, particularly in community-trained models. We introduce LanPaint, a training-free method tailored for widely adopted ODE-based samplers, which leverages Langevin dynamics to perform exact conditional inference, enabling precise visually coherent inpainting. LanPaint addresses two key challenges Langevin-based inpainting: (1) the risk of local likelihood maxima trapping (2) slow convergence. By proposing guided score function fast-converging framework, achieves high-fidelity results very few iterations. Experiments demonstrate that outperforms existing techniques, outperforming challenging tasks such as outpainting with Stable Diffusion.