VideoFlexTok: variable-length coarse-to-fine video tokenization
Original title:VideoFlexTok: Flexible-Length Coarse-to-Fine Video Tokenization
AI summary
VideoFlexTok proposes a variable-length token sequence video representation with a coarse-to-fine structure: the first token captures abstract information such as semantics and motion, subsequent tokens add fine details, and a generative flow decoder supports video reconstruction with any number of tokens. Compared with conventional 3D grid tokenization, this structure allows the token count to be adjusted for downstream needs and encodes longer videos under the same budget. In class- and text-to-video generation tasks, VideoFlexTok achieves comparable generation quality (gFVD and ViCLIP Score) with 1.1B parameters (one-fifth of 5.2B). Training a text-to-video model handling 10-second, 81-frame video requires only 672 tokens, 8 times fewer than an equivalent 3D grid tokenizer.
Key Takeaways
- VideoFlexTok proposes a variable-length token sequence video representation with a coarse-to-fine…
- Compared with conventional 3D grid tokenization,this structure allows the token count to be…
- In class- and text-to-video generation tasks,VideoFlexTok achieves comparable generation quality…
- 1B parameters (one-fifth of 5.
- Training a text-to-video model handling 10-second,81-frame video requires only 672 tokens,8 times…
Quick Facts
- Published
- Updated
- Category
- Papers & Research
- AI score
- 56
把视频 tokenization 从固定网格改成变长 coarse-to-fine,训练效率提升明显,还能做更长的视频。研究角度挺漂亮,但离产品落地还有距离,做视频生成的可以追一下。
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: VideoFlexTok: variable-length coarse-to-fine video tokenization. July 2, 2026. https://new.nat6.net/en/news/cmr3rd0l901busl7l11krygj3/