Research map / Video / Video generation
VideoFlexTok: Flexible-Length Coarse-to-Fine Video Tokenization
Andrei Atanov, Jesse Allardice, Roman Bachmann, Oğuzhan Fatih Kar, R Devon Hjelm, David Griffiths, Peter Fu, Afshin Dehghan, Amir Zamir
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 |
|---|---|---|---|
| use VideoFlexTok with 160 tokens instead of 1024-1280 token grid tokenizers | class-to-video AR model · Kinetics-600 | gFVD: 87.5 -> 80.0 | Improved |
| use VideoFlexTok vs Omnitokenizer | class-to-video AR model · Kinetics-600 | class alignment score: 0.858 -> 0.833 | Hurt |
| use VideoFlexTok vs LARP for reconstruction | tokenizer · Kinetics-600 | rFVD: 42.1 -> 48.7 | No clear effect |
| generate fewer VideoFlexTok tokens per frame with the AR model | class-to-video AR model · Kinetics-600 | model size / training tokens for comparable performance: 5-10x smaller model or 5-10x fewer training tokens | Improved |
| use VideoFlexTok with adjusted token count vs 3D grid tokenizer | text-to-video AR model · Panda70M | compute for comparable gFVD/ViCLIP: order of magnitude less compute | Improved |
| train on 32 tokens per frame for 10-second videos | text-to-video AR model · Panda70M 81-frame | token count per video: 5376 -> 672 | Improved |
| replace time-causal 2D registers with flat 1D registers (ablation) | VideoFlexTok · Kinetics-600 | rFVD: 69.9 -> 48.9 | Improved |
| replace time-causal 2D registers with flat 1D registers (ablation) | VideoFlexTok · Kinetics-600 class-to-video | gFVD: 287.6 -> 352.1 | Hurt |
| use full self-attention instead of time-causal decoder (ablation) | VideoFlexTok (32 tokens) · Kinetics-600 | rFVD: 80.9 -> 58.3 | Improved |
| use full self-attention instead of time-causal decoder (ablation) | VideoFlexTok (32 tokens) · Kinetics-600 class-to-video | gFVD: 175.1 -> 211.5 | Hurt |
Closest papers by meaning
- Efficient Long Video Tokenization via Coordinate-based Patch Reconstruction (CVPR 2025 · similarity 0.91)
- Video-LaVIT: Unified Video-Language Pre-training with Decoupled Visual-Motional Tokenization (ICML 2024 · similarity 0.89)
- Language Model Beats Diffusion - Tokenizer is key to visual generation (ICLR 2024 · similarity 0.89)
- VQToken: Neural Discrete Token Representation Learning for Extreme Token Reduction in Video Large Language Models (NeurIPS 2025 · similarity 0.89)
- When Worse is Better: Navigating the Compression Generation Trade-off In Visual Tokenization (NeurIPS 2025 · similarity 0.89)
- OmniTokenizer: A Joint Image-Video Tokenizer for Visual Generation (NeurIPS 2024 · 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.
