Physics-Grounded Reflection Simulation and Diffusion-Based Video Dereflection Framework Introduced

A novel closed-loop system for removing reflections from videos has been introduced in a paper available on arXiv (ID: 2608.11562). This system integrates physics-based reflection simulation, diffusion-driven video dereflection, and evaluation benchmarks. The S2R-Synthesis pipeline creates pairs of reflected and reflection-free videos by applying physics-informed augmentation in the structure space and utilizing a trained video diffusion renderer to produce realistic reflected visuals. Key effects related to glass, such as roughness blur, ghosting from thickness, and variations in reflectance, are modeled in the augmentation. The authors present S2R-Removal, the first specialized video dereflection model, addressing the often-overlooked challenge of video reflection removal, which lacks sufficient paired video data and coherent removal models. The framework seeks to enhance visual quality and support downstream vision tasks by eliminating reflections from glass-captured videos. This paper is marked as a cross-type submission on arXiv.