Research map / Video / Video-language understanding
SweetTok: Semantic-Aware Spatial-Temporal Tokenizer for Compact Video Discretization
Zhentao Tan, Ben Xue, Jian Jia, Junhao Wang, Wencai Ye, Shaoyun Shi, Mingjie Sun, Wenjin Wu, Quan Chen, Peng Jiang
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 |
|---|---|---|---|
| decouple spatial and temporal queries (DQAE) instead of vanilla query-based compression (ablation) | SweetTok · UCF-101 reconstruction | rFVD: 35.15 -> 20.46 | Improved |
| use decoupled queries instead of naive 1D linear-interpolation downsampling to 1280 tokens (ablation) | SweetTok · UCF-101 reconstruction | rFVD: 227.65 -> 20.46 | Improved |
| add vanilla language codebook (ablation) | SweetTok · UCF-101 reconstruction | rFVD: 29.45 -> 24.80 | Improved |
| make the language codebook motion-enhanced (verbs/adverbs for temporal queries) (ablation) | SweetTok · UCF-101 reconstruction | rFVD: 24.80 -> 20.46 | Improved |
| replace CLIP embeddings with larger Qwen embeddings in the MLC (ablation) | SweetTok · UCF-101 reconstruction | rFVD: 20.46 -> 20.12 | No clear effect |
| remove token compression (5,120 tokens instead of 1,280) | SweetTok · UCF-101 reconstruction | rFVD: 10.74 on UCF-101 and 7.51 on K-600 | Improved |
| scale the AR generator from 650M to 1.9B parameters | SweetTok · UCF-101 generation | gFVD: 84 -> 65 | Improved |
| use SweetTok tokens instead of OmniTok tokens for generation | VideoGPT AR generator · UCF-101 generation | gFVD: 191 -> 84 | Improved |
| use SweetTok tokens instead of LARP tokens for generation | AR generator · UCF-101 generation | gFVD: 99 -> 84 | Improved |
| fine-tune SweetTok spatial branch on ImageNet with 1,024 tokens | SweetTok · ImageNet reconstruction | rFID: 0.59 -> 0.37 | Improved |
Later papers that compared with this method
1 later accepted paper reports SweetTok as a baseline in their results tables; 0 of them report a better number on most of the results they share with it.
- PyraTok: Language-Aligned Pyramidal Tokenizer for Video Understanding and Generation (CVPR 2026 · better on 1 of 2 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- VideoFlexTok: Flexible-Length Coarse-to-Fine Video Tokenization (ICML 2026 · similarity 0.90)
- TokenLearner: Adaptive Space-Time Tokenization for Videos (NeurIPS 2021 · similarity 0.90)
- PVC: Progressive Visual Token Compression for Unified Image and Video Processing in Large Vision-Language Models (CVPR 2025 · similarity 0.89)
- OmniTokenizer: A Joint Image-Video Tokenizer for Visual Generation (NeurIPS 2024 · similarity 0.88)
- Efficient Long Video Tokenization via Coordinate-based Patch Reconstruction (CVPR 2025 · similarity 0.88)
- TokenFlow: Unified Image Tokenizer for Multimodal Understanding and Generation (CVPR 2025 · similarity 0.87)
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.
