FlexTok: Resampling Images into 1D Token Sequences of Flexible Length

Roman Bachmann, Jesse Allardice, David Mizrahi, Enrico Fini, Oğuzhan Fatih Kar, Elmira Amirloo, Alaaeldin El-Nouby, Amir Zamir, Afshin Dehghan
Proceedings of the 42nd International Conference on Machine Learning, PMLR 267:2241-2292, 2025.

Abstract

We introduce FlexTok, a tokenizer that projects 2D images into variable-length, ordered 1D token sequences. For example, a 256x256 image can be resampled into anywhere from 1 to 256 discrete tokens, hierarchically and semantically compressing its information. By training a rectified flow model as the decoder and using nested dropout, FlexTok produces plausible reconstructions regardless of the chosen token sequence length. We evaluate our approach in an autoregressive generation setting using a simple GPT-style Transformer. On ImageNet, this approach achieves an FID$<$2 across 8 to 128 tokens, outperforming TiTok and matching state-of-the-art methods with far fewer tokens. We further extend the model to support to text-conditioned image generation and examine how FlexTok relates to traditional 2D tokenization. A key finding is that FlexTok enables next-token prediction to describe images in a coarse-to-fine "visual vocabulary", and that the number of tokens to generate depends on the complexity of the generation task.

Cite this Paper

BibTeX
@InProceedings{pmlr-v267-bachmann25a, title = {{F}lex{T}ok: Resampling Images into 1{D} Token Sequences of Flexible Length}, author = {Bachmann, Roman and Allardice, Jesse and Mizrahi, David and Fini, Enrico and Kar, O\u{g}uzhan Fatih and Amirloo, Elmira and El-Nouby, Alaaeldin and Zamir, Amir and Dehghan, Afshin}, booktitle = {Proceedings of the 42nd International Conference on Machine Learning}, pages = {2241--2292}, year = {2025}, editor = {Singh, Aarti and Fazel, Maryam and Hsu, Daniel and Lacoste-Julien, Simon and Berkenkamp, Felix and Maharaj, Tegan and Wagstaff, Kiri and Zhu, Jerry}, volume = {267}, series = {Proceedings of Machine Learning Research}, month = {13--19 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v267/main/assets/bachmann25a/bachmann25a.pdf}, url = {https://proceedings.mlr.press/v267/bachmann25a.html}, abstract = {We introduce FlexTok, a tokenizer that projects 2D images into variable-length, ordered 1D token sequences. For example, a 256x256 image can be resampled into anywhere from 1 to 256 discrete tokens, hierarchically and semantically compressing its information. By training a rectified flow model as the decoder and using nested dropout, FlexTok produces plausible reconstructions regardless of the chosen token sequence length. We evaluate our approach in an autoregressive generation setting using a simple GPT-style Transformer. On ImageNet, this approach achieves an FID$<$2 across 8 to 128 tokens, outperforming TiTok and matching state-of-the-art methods with far fewer tokens. We further extend the model to support to text-conditioned image generation and examine how FlexTok relates to traditional 2D tokenization. A key finding is that FlexTok enables next-token prediction to describe images in a coarse-to-fine "visual vocabulary", and that the number of tokens to generate depends on the complexity of the generation task.} }
Endnote
%0 Conference Paper %T FlexTok: Resampling Images into 1D Token Sequences of Flexible Length %A Roman Bachmann %A Jesse Allardice %A David Mizrahi %A Enrico Fini %A Oğuzhan Fatih Kar %A Elmira Amirloo %A Alaaeldin El-Nouby %A Amir Zamir %A Afshin Dehghan %B Proceedings of the 42nd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2025 %E Aarti Singh %E Maryam Fazel %E Daniel Hsu %E Simon Lacoste-Julien %E Felix Berkenkamp %E Tegan Maharaj %E Kiri Wagstaff %E Jerry Zhu %F pmlr-v267-bachmann25a %I PMLR %P 2241--2292 %U https://proceedings.mlr.press/v267/bachmann25a.html %V 267 %X We introduce FlexTok, a tokenizer that projects 2D images into variable-length, ordered 1D token sequences. For example, a 256x256 image can be resampled into anywhere from 1 to 256 discrete tokens, hierarchically and semantically compressing its information. By training a rectified flow model as the decoder and using nested dropout, FlexTok produces plausible reconstructions regardless of the chosen token sequence length. We evaluate our approach in an autoregressive generation setting using a simple GPT-style Transformer. On ImageNet, this approach achieves an FID$<$2 across 8 to 128 tokens, outperforming TiTok and matching state-of-the-art methods with far fewer tokens. We further extend the model to support to text-conditioned image generation and examine how FlexTok relates to traditional 2D tokenization. A key finding is that FlexTok enables next-token prediction to describe images in a coarse-to-fine "visual vocabulary", and that the number of tokens to generate depends on the complexity of the generation task.
APA
Bachmann, R., Allardice, J., Mizrahi, D., Fini, E., Kar, O.F., Amirloo, E., El-Nouby, A., Zamir, A. & Dehghan, A.. (2025). FlexTok: Resampling Images into 1D Token Sequences of Flexible Length. Proceedings of the 42nd International Conference on Machine Learning, in Proceedings of Machine Learning Research 267:2241-2292 Available from https://proceedings.mlr.press/v267/bachmann25a.html.

Related Material