[论文] DUO-VSR: Dual-Stream Distillation for One-Step Video Super-Resolution

论文概要

研究领域: CV 作者: Zhengyao Lv, Menghan Xia, Xintao Wang, Kwan-Yee K. Wong 发布时间: 2026-03-23 arXiv: 2603.22271

中文摘要

基于扩散的视频超分辨率（VSR）最近取得了显著的保真度，但仍面临高昂的采样成本。虽然分布匹配蒸馏（DMD）可以将扩散模型加速到单步生成，但直接将其应用于VSR往往会导致训练不稳定，同时监督信号退化且不足。为解决这些问题，我们提出了DUO-VSR，这是一个基于双流蒸馏策略的三阶段框架，将分布匹配与对抗监督统一于单步VSR中。首先，采用渐进引导蒸馏初始化，通过保持轨迹的蒸馏来稳定后续训练。其次，双流蒸馏联合优化DMD和Real-Fake Score Feature GAN（RFS-GAN）流，后者利用来自真实和虚假分数模型的判别特征提供互补的对抗监督。最后，偏好引导精修阶段进一步将学生模型与感知质量偏好对齐。

原文摘要

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, we 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. Firstly, a Progressive Guided Distillation Initialization is employed to stabilize subsequent training through trajectory-preserving distillation. Next, the Dual-Stream Distillation jointly optimizes the DMD and Real-Fake Score Feature GAN (RF...

自动采集于 2026-03-25

#论文 #arXiv #CV #小凯