DOVE Logo DOVE

Efficient One-Step Diffusion Model for Real-World Video Super-Resolution

Zheng Chen1*, Zichen Zou1*, Kewei Zhang1, Xiongfei Su2, Xin Yuan2, Yong Guo3, Yulun Zhang1†
1Shanghai Jiao Tong University, 2Westlake University, 3Huawei Consumer Business Group
*Indicates Equal contribution, †Indicates Corresponding author
arXiv Supplementary Code
Dataset Pretrained Model

Real-World Videos (upscale ×4)

Abstract

Diffusion models have demonstrated promising performance in real-world video super-resolution (VSR). However, the dozens of sampling steps they require, make inference extremely slow. Sampling acceleration techniques, particularly single-step, provide a potential solution. Nonetheless, achieving one step in VSR remains challenging, due to the high training overhead on video data and stringent fidelity demands. To tackle the above issues, we propose DOVE, an efficient one-step diffusion model for real-world VSR. DOVE is obtained by fine-tuning a pretrained video diffusion model (i.e., CogVideoX). To effectively train DOVE, we introduce the latent–pixel training strategy. The strategy employs a two-stage scheme to gradually adapt the model to the video super-resolution task. Meanwhile, we design a video processing pipeline to construct a high-quality dataset tailored for VSR, termed HQ-VSR. Fine-tuning on this dataset further enhances the restoration capability of DOVE. Extensive experiments show that DOVE exhibits comparable or superior performance to multi-step diffusion-based VSR methods. It also offers outstanding inference efficiency, achieving up to a 28× speed-up over existing methods such as MGLD-VSR.

Method

DOVE Training Strategy
Overview of the framework and training strategy of DOVE

Our method performs one-step sampling to reconstruct high-resolution (HR) videos (xsr) from low-resolution (LR) inputs (xlr). To enable effective training, we adopt a two-stage latent-pixel training strategy. Stage-1 (latent-space): Minimize the difference between the predicted and HR latents, allowing the model to learn one-step LR-to-HR mapping. Stage-2 (pixel-space): Improve detail generation using mixed image/video training.

HQ-VSR

HQ-VSR Dataset
Overview of the HQ-VSR dataset construction pipeline

We design a systematic video processing pipeline to curate HQ-VSR, a high-quality dataset for video super-resolution. Starting from OpenVid-1M, we extract 2,055 diverse and high-quality videos suitable for VSR.

Results

Quantitative Results (click to expand)
  • Quantitative Results (Tab. 2 of the main paper)

  • Complexity Comparison (Tab. 2 of the supplementary material)

Qualitative Results (click to expand)
  • Efficiency and Performance Comparison (Fig. 1 of the main paper)

  • Visual Results (Fig. 4 of the main paper)

    More Results (click to expand)
    • Multi-frame comparison on the synthetic dataset (UDM10) (Fig. 3 of the supplementary material)

    • Multi-frame comparison on the real-world dataset (RealVSR) (Fig. 4 of the supplementary material)

    • More results on synthetic (UDM10, SPMCS, and YouHQ40) datasets (Fig. 5 of the supplementary material)

    • More results on real-world (RealVSR, MVSR4x, and VideoLQ) datasets (Fig. 6 and 7 of the supplementary material)

Comparison with SOTA

  • Slide 1
  • Slide 2
  • Slide 3
  • Slide 6
  • Slide 10
LR
Upscale-A-Video
STAR
SeedVR
SeedVR2
DOVE (ours)

BibTeX

  @inproceedings{chen2025dove,
    title={DOVE: Efficient One-Step Diffusion Model for Real-World Video Super-Resolution},
    author={Chen, Zheng and Zou, Zichen and Zhang, Kewei and Su, Xiongfei and Yuan, Xin and Guo, Yong and Zhang, Yulun},
    booktitle={NeurIPS},
    year={2025}
  }