Reproduce “Generative Video Compression with Adaptive Score Distillation” 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
Naifu Xue, Zhaoyang Jia, Haosen Li, Zihan Zheng, Jiahao Li, Bin Li, Xiaoyi Zhang, Qi Meng and 2 more
Researchers developed GenVC, a Generative Video Codec that employs a pixel-domain video diffusion model optimized specifically for compression and accelerated by one-step Adaptive Score Distillation. The system achieves high perceptual quality at ultra-low bitrates with efficient decoding, outperforming prior generative codecs.
Key insights
- A video diffusion model trained from scratch, specifically for compression and jointly with the codec, can yield superior perceptual quality at constrained bitrates compared to adapting general-purpose pre-trained models.
- A global-to-local hierarchical architecture effectively manages the spatio-temporal complexity of pixel-domain video diffusion, enabling high-resolution modeling without prohibitive computational costs.
- Direct application of Distribution Matching Distillation (DMD) for video can suffer from 'teacher-side guidance failure,' where the frozen teacher provides misleading updates that result in motion-stalled student reconstructions.
- Adaptive Score Distillation, by selectively weighting DMD updates and integrating source-referenced distortion losses, is crucial for stable and effective one-step video distillation that preserves motion coherence.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.