LanPaint: Training-Free Diffusion Inpainting with Asymptotically Exact and Fast Conditional Sampling

Authors: Candi Zheng, Yuan Lan, Yang Wang

TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments demonstrate that our approach achieves superior performance with precise partial conditioning and visually coherent inpainting across diverse tasks. Experiments confirm that Lan Paint outperforms existing training-free approaches, delivering high-quality inpainting and outpainting results for both pixel-space and latent-space models. We evaluate the inpainting performance of Lan Paint on the Celeb A-HQ-256 (Liu et al., 2015) and Image Net256 (Deng et al., 2009) datasets, leveraging pre-trained latent (Rombach et al., 2021) and pixel space (Dhariwal & Nichol, 2021) diffusion models, respectively. The experiments assess reconstruction quality across various mask geometries, including box, half, checkerboard, and outpainting. Perceptual fidelity is quantified through LPIPS (Zhang et al., 2018) and FID metrics, calculated on 1,000 validation images per dataset. Results are presented in Tables 1 and 2.
Researcher Affiliation Academia Candi ZHENG EMAIL Department of Mathematics, Hong Kong University of Science and Technology, Yuan LAN EMAIL Independent Researcher, Yang Wang EMAIL Department of Mathematics, Hong Kong University of Science and Technology
Pseudocode Yes Algorithm 4: Lan Paint, Variance Perserving Notation
Open Source Code Yes Code is available on https://github.com/scraed/Lan Paint.
Open Datasets Yes We evaluate the inpainting performance of Lan Paint on the Celeb A-HQ-256 (Liu et al., 2015) and Image Net256 (Deng et al., 2009) datasets
Dataset Splits Yes perceptual fidelity is quantified through LPIPS (Zhang et al., 2018) and FID metrics, calculated on 1,000 validation images per dataset.
Hardware Specification Yes Evaluations were conducted on a single RTX 3090.
Software Dependencies No The paper mentions "We implement Lan Paint using the diffusers package, following Algorithm 4." and "Images are generated through Comfy UI (Comfy Org, 2025) with Euler sampler (Karras et al., 2022) (30 steps)." However, it does not provide specific version numbers for the 'diffusers package' or 'Comfy UI', which are required for a reproducible description of ancillary software.
Experiment Setup Yes Hyperparameters are configured as follows: γ = 15, α = 0., and λ = 8 for all image inpainting tasks, drawing loosely from the insights gained through sensitivity analysis in Fig.8. The notation Lan Paint-5 and Lan Paint-10 denotes N = 5 and N = 10 sampling steps, respectively. The step size η is set to 0.15 for both Celeb-A and Image Net.