ElasticTok: Adaptive Tokenization for Image and Video

Wilson Yan, Volodymyr Mnih, Aleksandra Faust, Matei Zaharia, Pieter Abbeel, Hao Liu · ICLR 2025 · 2025

arXiv Code

Status

To read

Topics

budgeted-inference, other

Feeds

P2 · NeurIPS 2027

What they did

Encodes each image or video frame into a variable number of tokens, conditioned on earlier frames, by training with a random number of tokens dropped from the end of each frame.

My note

Trains representations that support variable token budgets according to content complexity.

Code

Repository

LargeWorldModel/ElasticTok

Framework

—

Licence

Apache-2.0

My status

Not tried

BibTeX

@inproceedings{yan2025-elastictok,
  title = {{ElasticTok: Adaptive Tokenization for Image and Video}},
  author = {Wilson Yan and Volodymyr Mnih and Aleksandra Faust and Matei Zaharia and Pieter Abbeel and Hao Liu},
  booktitle = {ICLR 2025},
  year = {2025},
  eprint = {2410.08368},
  archivePrefix = {arXiv}
}