Summary
DUO-VSR is a new framework for diffusion-based video super-resolution (VSR) that reduces the prohibitive sampling cost of diffusion models to a single generation step. Directly applying distribution matching distillation (DMD) to VSR often causes training instability and degraded, insufficient supervision. To address this, the authors propose a three-stage framework built on a Dual-Stream Distillation strategy that unifies distribution matching and adversarial supervision. Stage one uses Progressive Guided Distillation Initialization, a trajectory-preserving distillation that stabilizes subsequent training. Stage two jointly optimizes a DMD stream and a Real-Fake Score Feature GAN (RFS-GAN) stream, where the GAN leverages discriminative features from both real and fake score models to provide complementary adversarial supervision. Stage three is a preference-guided refinement that aligns the student model with perceptual quality preferences. Authored by Zhengyao Lv, Menghan Xia, Xintao Wang, and Kwan-Yee K. Wong, the paper is available on arXiv (2603.22271).
Overview
Field: Computer Vision
Authors: Zhengyao Lv, Menghan Xia, Xintao Wang, Kwan-Yee K. Wong
Published: 2026-03-23
arXiv: 2603.22271
Abstract
Diffusion-based video super-resolution (VSR) has recently achieved remarkable fidelity but still suffers from prohibitive sampling costs. While distribution matching distillation (DMD) can accelerate diffusion models toward one-step generation, directly applying it to VSR often results in training instability alongside degraded and insufficient supervision. To address these issues, the authors propose DUO-VSR, a three-stage framework built upon a Dual-Stream Distillation strategy that unifies distribution matching and adversarial supervision for one-step VSR:
1. Progressive Guided Distillation Initialization: stabilizes subsequent training through trajectory-preserving distillation.
2. Dual-Stream Distillation: jointly optimizes the DMD stream and a Real-Fake Score Feature GAN (RFS-GAN) stream, which exploits discriminative features from both real and fake score models to provide complementary adversarial supervision.
3. Preference-Guided Refinement: further aligns the student model with perceptual quality preferences.
Links
- Paper: https://arxiv.org/abs/2603.22271
---
*Auto-collected on 2026-03-25 from zhichai.net.*
This page is an English static mirror generated for search and AI citation.
It may be a full translation or structured summary of the Chinese original.
Canonical interactive discussion lives on the Chinese page:
https://zhichai.net/topic/177169035