From Synthesis to Removal: Physics-Grounded Reflection Simulation and Diffusion-Based Video Dereflection
The paper claims the first diffusion-based model and benchmark for removing glass reflections from video.
The authors build a closed-loop system: synthesize paired reflected and clean videos, train a removal model on them, then evaluate on a dedicated benchmark. Their synthesis pipeline models glass effects such as blur from roughness, ghosting from thickness, and reflectance changes. The removal model adapts a pretrained video diffusion prior and recovers the clean transmission in one denoising step. They report state-of-the-art results on their benchmark and public image benchmarks, with faster inference than non-diffusion baselines. Source: HF Daily Papers' note.
The authors build a closed-loop system: synthesize paired reflected and clean videos, train a removal model on them, then evaluate on a dedicated benchmark. Their synthesis pipeline models glass effects such as blur from roughness, ghosting from thickness, and reflectance changes. The removal model adapts a pretrained video diffusion prior and recovers the clean transmission in one denoising step. They report state-of-the-art results on their benchmark and public image benchmarks, with faster inference than non-diffusion baselines. Source: HF Daily Papers' note.
score 4