Research map / Video / Video generation
TokensGen: Harnessing Condensed Tokens for Long Video Generation
Wenqi Ouyang, Zeqi Xiao, Danni Yang, Yifan Zhou, Shuai Yang, Lei Yang, Jianlou Si, Xingang Pan
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 condensed tokens 4x8x12 instead of 4x5x7 (ablation) | TokensGen To2V · To2V reconstruction | MSE: 2.84e-2 -> 1.24e-2 | Improved |
| increase token spatial size from 5x7 to 8x12 at 4 frames vs 13x5x7 (ablation) | TokensGen To2V · To2V reconstruction | MSE: 1.54e-2 -> 1.24e-2 | Improved |
| add PCA channel projection to condensed tokens (ablation) | TokensGen To2V · To2V reconstruction | MSE / SSIM: 1.24e-2 -> 1.24e-2; SSIM 0.62 -> 0.62 | No clear effect |
| use condensed tokens instead of super-resolution conditioning (ablation) | TokensGen To2V · To2V reconstruction | SSIM: 0.60 -> 0.62; MSE 1.12e-2 -> 1.24e-2 | Mixed |
| use condensed tokens instead of super-resolution conditioning (ablation) | TokensGen To2V · To2V reconstruction | temporal flickering: 97.10 -> 97.58 | Improved |
| add FIFO-Diffusion to clip-wise generation (ablation) | TokensGen · long video | visual consistency | Improved |
| add adaptive padding to FIFO (ablation) | TokensGen · long video | initial-frame quality | Improved |
| use TokensGen instead of AP-FIFO+CogVideoX | long video generation · Human study | Text-Visual Alignment: 24.31% -> 75.69% | Improved |
| use TokensGen instead of AP-FIFO+CogVideoX | long video generation · Human study | Motion & Content Consistency: 22.57% -> 75.74% | Improved |
| use TokensGen instead of AP-FIFO+CogVideoX | long video generation · VBench | Subject Consistency: 86.22 -> 84.57 | Hurt |
| use TokensGen instead of AP-FIFO+CogVideoX | long video generation · VBench | Motion Smoothness: 97.48 -> 98.08 | Improved |
| use TokensGen instead of AP-FIFO+CogVideoX | long video generation · VBench | Dynamic Degree: 78.57 -> 78.95 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DiTCtrl — DiTCtrl: Exploring Attention Control in Multi-Modal Diffusion Transformer for Tuning-Free Multi-Prompt Longer Video Generation (CVPR 2025) · better on 4 of 4 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
- SEINE: Short-to-Long Video Diffusion Model for Generative Transition and Prediction (ICLR 2024 · similarity 0.90)
- StreamingT2V: Consistent, Dynamic, and Extendable Long Video Generation from Text (CVPR 2025 · similarity 0.90)
- TokenFlow: Consistent Diffusion Features for Consistent Video Editing (ICLR 2024 · similarity 0.89)
- LinGen: Towards High-Resolution Minute-Length Text-to-Video Generation with Linear Computational Complexity (CVPR 2025 · similarity 0.89)
- Language Model Beats Diffusion - Tokenizer is key to visual generation (ICLR 2024 · similarity 0.89)
- Long Video Diffusion Generation with Segmented Cross-Attention and Content-Rich Video Data Curation (CVPR 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.
