FeaTok replaces the usual pixel-reconstruction objective for discrete image tokenizers with reconstruction of features from pretrained SigLIP, SigLIP2, DINOv2, or DINOv3 encoders.
Feature-trained image tokens improve generation quality at lower sampling cost
FeaTok learns discrete tokens by reconstructing pretrained visual features, outperforming pixel-trained tokenizers across class-conditional and text-to-image evaluations.
Top University
Zhang Xiaofeng · Samuel Lavoie · Nicolas Beltran-Velez · Jianhao Yuan · Felix Friedrich · Jakob Verbeek · +3 more
Mila - Québec AI Institute · Université de Montréal · FAIR at Meta · Columbia University · University of Oxford
Research Digest··3 min read
Thread:Visual Token Efficiency
Zhang Xiaofeng and colleagues separate token learning from pixel rendering: the tokenizer is trained to reconstruct semantic features, while a detached decoder converts tokens into images without letting pixel losses reshape the representation.
Why this paper
From Mila - Québec AI Institute and 6 others · Part of Visual Token Efficiency, now 6 papers
In one line
Training discrete image tokens with feature reconstruction instead of pixel reconstruction improves generative quality and efficiency.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§