Video Scratch Removal Method Based on Guided Diffusion Generation With Locally Constrained Generation

ICLR 2026 Conference Submission15828 Authors

Keywords: Video Scratch Removal

Abstract: This paper proposes a video scratch removal method based on guided diffusion with locally constrained generation. By estimating optical flow between adjacent frames and integrating visual features, we construct a guidance map that encodes both temporal and spatial information. A relaxed mask is computed using a mean filter, which helps mitigate motion estimation errors while effectively distinguishing potential scratch regions.To further reduce motion-induced blurring during the restoration process, the proposed method leverages the guidance map---fused from temporal and spatial cues of neighboring frames---as a conditioning input to the diffusion model for restoring the current frame. In addition, the relaxed mask is used to constrain the generation to local regions, allowing uncorrupted areas to retain their original content.Experimental results demonstrate that our approach not only significantly improves restoration quality but also effectively reduces the likelihood of missing scratches.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 15828