Research map / Video / Video generation
EVATok: Adaptive Length Video Tokenization for Efficient Visual Autoregressive Generation
Tianwei Xiong, Jun Hao Liew, Zilong Huang, Zhijie Lin, Jiashi Feng, Xihui Liu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| assign tokens with the router instead of uniform 1024 tokens (no discriminator) | final tokenizer · WebVid val | rFVD: 63 -> 33 with 1024 -> 721 tokens | Improved |
| assign tokens with the router instead of uniform 1024 tokens (+ VideoMAE discriminator) | final tokenizer · WebVid val | rFVD: 13 -> 9.2 with 1024 -> 721 tokens | Improved |
| assign tokens with the router instead of uniform on UCF | final tokenizer + GPT-B · UCF-101 | gFVD: 98 -> 96 with 1024 -> 740 generated tokens | Improved |
| assign tokens with the router instead of uniform on UCF | final tokenizer · UCF-101 | LPIPS: 0.1303 -> 0.1212 | Improved |
| train a final tokenizer from scratch instead of reusing the proxy tokenizer (uniform) (ablation) | uniform tokenizer · WebVid val | rFVD: 73 -> 63 | Improved |
| train a final tokenizer from scratch instead of reusing the proxy tokenizer (router) (ablation) | router tokenizer · WebVid val | rFVD: 50 -> 33 | Improved |
| add VideoMAE discriminator (ablation) | final recipe · UCF-101 | rFVD: 65 -> 13 | Improved |
| add VideoMAE discriminator (ablation) | final recipe · UCF-101 | gFVD: 155 -> 98 | Improved |
| add V-JEPA2 representation alignment (ablation) | final recipe · UCF-101 | gFVD: 144 -> 98 | Improved |
| add both VideoMAE discriminator and V-JEPA2 alignment (ablation) | final recipe · UCF-101 | rFVD: 80 -> 13 | Improved |
| add VideoMAE discriminator | uniform final tokenizer · WebVid val | PSNR: 27.77 -> 26.68 | Hurt |
| use max-proxy-reward assignment or router instead of uniform assignment | proxy tokenizer · WebVid and UCF | rFVD: 56% token savings on WebVid and 42% on UCF | Improved |
Closest papers by meaning
- LARP: Tokenizing Videos with a Learned Autoregressive Generative Prior (ICLR 2025 · similarity 0.91)
- Parallelized Autoregressive Visual Generation (CVPR 2025 · similarity 0.90)
- Vision Foundation Models as Effective Visual Tokenizers for Autoregressive Generation (NeurIPS 2025 · similarity 0.89)
- Learnings from Scaling Visual Tokenizers for Reconstruction and Generation (ICML 2025 · similarity 0.89)
- Efficient Long Video Tokenization via Coordinate-based Patch Reconstruction (CVPR 2025 · similarity 0.89)
- When Worse is Better: Navigating the Compression Generation Trade-off In Visual Tokenization (NeurIPS 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
