Researchers have introduced a closed-loop framework designed to address the challenges of video reflection removal, a task often hindered by the lack of paired datasets and temporal coherence. The proposed solution begins with S2R-Synthesis, a pipeline that generates paired reflected and reflection-free video data. This synthesis method employs physics-grounded augmentation in the structure space and utilizes a trained video diffusion renderer to simulate complex glass-related phenomena, including thickness-induced ghosting, roughness-induced blur, and varying reflectance levels. Leveraging this synthesized data, the authors developed S2R-Removal, the inaugural diffusion-based model for video dereflection. This model adapts a pretrained video diffusion prior through reflection-aware latent adaptation and incorporates a one-step pixel-geometric refinement technique, allowing it to recover clean transmission signals in a single denoising step. To facilitate future research, the authors established S2R-Bench, a new benchmark supporting both full-reference evaluation and human perceptual assessment. Experimental results on S2R-Bench and existing image benchmarks indicate that the S2R-Removal model achieves state-of-the-art performance while maintaining faster inference speeds compared to non-diffusion baselines. This framework effectively bridges the gap between physics-based simulation and generative video restoration.