Reproduce “Autoregressive Video Generation beyond Next Frames Prediction” 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
Sucheng Ren, Chen Chen, Zhenbang Wang, Liangchen Song, Xiangxin Zhu, Alan Yuille, Yinfei Yang, Jiasen Lu
The VideoAR framework introduces a generalized autoregressive approach for video generation, leveraging spatiotemporal cubes as prediction units and a novel Symmetric Distribution Matching Distillation training. It achieves a VBench total score of 84.87, outperforming state-of-the-art diffusion models while generating videos significantly faster at 16.4 Frames Per Second and enabling zero-shot minute-long content.
Key insights
- Spatiotemporal cubes significantly improve video generation quality and temporal coherence compared to traditional frame-based or other prediction units in autoregressive models.
- An autoregressive architecture, when combined with a robust continuous latent space training objective like Symmetric DMD, can surpass state-of-the-art diffusion models in quality while maintaining much higher inference speeds.
- By designing prediction units that capture local appearance and motion jointly and employing a streaming autoregressive process with KV cache management, ultra-long videos can be generated coherently in a zero-shot manner.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.