Reproduce “Rethinking Video Deblurring with Wavelet-Aware Dynamic Transformer and Diffusion Model” 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
Chen Rao, Guangyuan Li, Zehua Lan, Jiakai Sun, Junsheng Luan, Wei Xing, Lei Zhao, Huaizhong Lin and 2 more
Researchers at Zhejiang University and Zhejiang Gongshang University introduce VD-Diff, a framework integrating a Wavelet-Aware Dynamic Transformer with a Diffusion Model operating in a compact latent space. It achieves new state-of-the-art performance on video deblurring benchmarks, including 35.55 dB PSNR on GoPro, by effectively recovering high-frequency details and maintaining spatio-temporal consistency.
Key insights
- Diffusion Models can be effectively leveraged to provide high-frequency prior information for video deblurring, successfully compensating for the detail-loss in regression-based approaches.
- Applying Diffusion Models in a highly compact latent space, rather than pixel space, enables efficient high-quality generation with significantly fewer iterative steps, making them practical for video tasks.
- A frequency-aware Transformer architecture, combined with bidirectional temporal propagation, is crucial for both robust low-frequency structure recovery and coherent spatio-temporal consistency in deblurred videos.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.