Reproduce “OS-DiffVSR: Towards One-step Latent Diffusion Model for High-detailed Real-world 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
Hanting Li, Huaao Tang, Jianhong Han, Tianxiong Zhou, Jiulong Cui, Haizhen Xie, Yan Chen, Jie Hu
OS-DiffVSR, developed by Huawei Noah's Ark Lab, introduces a one-step latent diffusion model for real-world video super-resolution, achieving superior perceptual quality and high inference efficiency. It reconstructs high-detailed videos from low-resolution inputs in a single sampling step, effectively addressing the traditional quality-efficiency trade-off in diffusion-based VSR.
Key insights
- Leveraging a foundation model like DINOv2 as a patch-level discriminator in the AFAT paradigm provides precise, local guidance for generating more realistic details in one-step diffusion VSR.
- The MFF module's design, which fuses temporal information in the latent space before the UNet and incorporates hard masking, effectively improves temporal consistency without excessive computational overhead.
- The Focal Mean Square Error (FMSE) loss, which assigns higher weights to spatial locations with poorer visual quality, directs the model to focus on difficult-to-reconstruct regions, leading to improved overall detail and realism.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.