LanPaint#

Training-Free Conditional Sampling for Inpainting and Local Image Editing#

LanPaint is a training-free partial conditional sampler for pretrained diffusion and rectified-flow models. It enables mask-constrained inpainting and local image editing without model fine-tuning or backpropagation.

LanPaint was introduced in the TMLR paper “LanPaint: Training-Free Diffusion Inpainting with Asymptotically Exact and Fast Conditional Sampling” by Candi Zheng, Yuan Lan, and Yang Wang.

Explore LanPaint#

What is LanPaint?#

LanPaint is a training-free partial conditional sampling method for diffusion and rectified-flow generative models.

Given a pretrained generative model and a spatial constraint such as a known image region or mask, LanPaint samples the unknown region while conditioning on the known region. It does this without training an additional inpainting model, fine-tuning the foundation model, or using backpropagation during inference.

The method uses carefully designed Langevin dynamics to perform fast, backpropagation-free Monte Carlo sampling and is compatible with ODE-based and rectified-flow models.

This makes LanPaint useful not only as an inpainting method, but more generally as a training-free mechanism for spatially constrained generation and local editing.

For the local-editing perspective, see LanPaint for Training-Free Local Editing.

For a technical explanation of partial conditional sampling, BiG Score, and Fast Langevin Dynamics, see LanPaint: Training-Free Partial Conditional Sampling with Langevin Dynamics.

Is LanPaint only an image inpainting method?#

No. Inpainting is the primary task studied in the original LanPaint paper, but the underlying method is partial conditional sampling.

This distinction matters because many local image-editing tasks can also be formulated as generation under spatial constraints: part of the image should remain fixed while another region is resampled or modified.

In practice, LanPaint can therefore serve as a training-free sampling layer for mask-constrained editing with pretrained generative models.

Can pretrained diffusion models perform local image editing without fine-tuning?#

LanPaint enables pretrained diffusion and rectified-flow models to perform mask-constrained local editing without additional fine-tuning or backpropagation.

The same conditional-sampling mechanism can support tasks such as:

• image inpainting
• outpainting and generative fill
• mask-guided local image editing
• object or region replacement
• localized content generation
• video inpainting and local video editing
• masked video and audio generation

The foundation model provides the generative prior, while LanPaint provides the training-free conditional sampling mechanism.

LanPaint for Training-Free Local Image Editing#

Many image-editing systems require task-specific training, adapters, inversion procedures, or optimization at inference time.

LanPaint follows a different approach: it operates directly on pretrained diffusion and rectified-flow models and performs partial conditional sampling at inference time.

This places LanPaint at the intersection of:

training-free image editing · local image editing · mask-guided generation · conditional sampling · diffusion models · rectified flow

For researchers working on training-free local editing, LanPaint can be viewed as a general sampling mechanism for preserving known regions while generating or modifying selected regions.

LanPaint and Rectified-Flow Models#

LanPaint is designed to work with ODE-based diffusion samplers and rectified-flow models, rather than depending only on stochastic DDPM sampling.

This is an important distinction from many earlier conditional-sampling and diffusion inverse-problem methods.

As rectified-flow and flow-based foundation models become increasingly common, LanPaint provides a training-free way to add partial conditioning and mask-constrained generation to pretrained models without retraining them.

What can LanPaint be used for?#

LanPaint has been implemented across multiple modern generative-model families and used for image, video, and multimodal masked generation.

Its central use case is simple:

keep one part of the generated sample constrained while allowing another part to be generated or edited.

This makes the same sampling mechanism useful across different foundation models and different spatially constrained generation tasks.

See the GitHub repository for the latest supported models, workflows, and examples.

How is LanPaint used beyond inpainting?#

Subsequent research has already begun using or classifying LanPaint outside the narrow setting of conventional image inpainting.

Why use LanPaint instead of training an inpainting or editing model?#

LanPaint is designed for cases where a strong pretrained generative model already exists and the goal is to obtain conditional generation capability without retraining that model.

Its key properties are:

Training-free. No inpainting-specific or editing-specific model training is required.

Backpropagation-free inference. LanPaint does not require gradient-based optimization through the generative model during sampling.

Model reuse. A pretrained generative prior can be reused for spatially constrained generation.

ODE and rectified-flow compatibility. LanPaint is designed for modern deterministic sampling and flow-model settings.

General conditional-sampling perspective. Inpainting is treated as an instance of partial conditional sampling rather than as a separate model architecture.

Research Resources#

Paper
LanPaint: Training-Free Diffusion Inpainting with Asymptotically Exact and Fast Conditional Sampling
Candi Zheng, Yuan Lan, Yang Wang
Transactions on Machine Learning Research, 2025

TMLR / OpenReview · arXiv

Implementations

Official ComfyUI implementation
Diffusers implementation
Hugging Face

Benchmark

LanPaintBench

Citation#

If LanPaint is useful in your research, please cite the TMLR paper:

1
@article{
2
zheng2025lanpaint,
3
title={LanPaint: Training-Free Diffusion Inpainting with Asymptotically Exact and Fast Conditional Sampling},
4
author={Candi Zheng and Yuan Lan and Yang Wang},
5
journal={Transactions on Machine Learning Research},
6
issn={2835-8856},
7
year={2025},
8
url={https://openreview.net/forum?id=JPC8JyOUSW}
9
}