Reproduce “DUO-VSR: Dual-Stream Distillation for One-Step Video Super-Resolution” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Zhengyao Lv, Menghan Xia, Xintao Wang, Kwan-Yee K. Wong

DUO-VSR introduces a dual-stream distillation framework for one-step video super-resolution, achieving high visual quality and strong temporal consistency with significantly accelerated inference. It speeds up video processing by up to 90x compared to multi-step methods and over 5x compared to other one-step approaches, while outperforming them on perceptual metrics and user studies.

Key insights

  • A multi-stage approach, starting with progressive guided distillation, provides a stable initialization for one-step VSR, mitigating large distribution shifts.
  • The joint optimization of distribution matching and adversarial supervision (from real videos) via a dual-stream strategy is crucial for overcoming supervision limitations and enhancing perceptual quality and temporal consistency in one-step VSR.
  • Careful design of the adversarial component, including using diffused samples, noise perturbation, and a stop-gradient mechanism, is essential for stabilizing GAN training and preventing interference with score model distribution tracking.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.