News Note on a New Diffusion-Based Method for Removing Reflections from Video

Researchers presented the work From Synthesis to Removal: Physics-Grounded Reflection Simulation and Diffusion-Based Video Dereflection, published on August 12, 2026. The authors describe a closed-loop framework of three components: S2R-Synthesis for generating paired data with and without reflections, S2R-Removal — the first diffusion model for removing reflections from video, and S2R-Bench — the first benchmark for evaluating the quality of such removal. The removal model adapts a pretrained video diffusion prior and restores a clean image in a single denoising step. The authors did not release the code.

What it means

The topic falls under the AI for video section but does not directly intersect with the tools in our reference. S2R-Removal is a video restoration model, not a generator. It takes video corrupted by reflections and cleans it, whereas LTX-Video, CogVideoX, and Pyramid Flow create new video from a text prompt or an image. There is an indirect connection: S2R-Synthesis uses a trained video diffusion renderer to generate realistic reflections, and S2R-Removal adapts a pretrained video diffusion prior. This shows that the diffusion architectures underlying the generators in our reference are also applicable to restoration tasks. However, none of our tools offers a reflection removal feature, and the authors of the new work do not disclose which specific video diffusion prior they adapted. There is no practical application yet for users of the video generators in our reference: no code, no weights, and the benchmark is not published. The work remains at the level of a research publication.