MixTok: Adaptive Visual Tokenization with Global 1D Tokens and Sparse 2D Refinement
Chenge Li, Yefei He, Xinyu Liu, Feng Chen, Hong Zhou, Bohan Zhuang
Zhejiang University · The Hong Kong University of Science and Technology · University of Adelaide
Published October 9, 2026 · on the feed October 12, 2026 · ▲ 13 · arXiv 2610.13181
The abstract
Efficient visual tokenization requires balancing representation compactness with the preservation of fine-grained image content. Conventional grid-based tokenizers allocate tokens uniformly across spatial locations despite non-uniform visual complexity, whereas compact 1D tokenizers aggregate global information but can struggle to preserve local details. We introduce MixTok, a hybrid visual tokenizer that combines a compact 1D global representation with sparse 2D tokens for local refinement. A content-adaptive router ranks candidate locations by refinement priority and determines the number of 2D tokens for each image under a target average token budget. At the highest-scoring locations, 2D token features replace the spatial features recovered from the 1D representation, and a shared decoder reconstructs the image from the resulting mixed features. On ImageNet 256 x 256, MixTok achieves highly competitive reconstruction and generation performance, with an rFID of 1.195 and a gFID of 1.95 at respective average budgets of approximately 128 and 131 content tokens per image.
The Brief
2:01A short podcast summary of the paper. Written and voiced by AI.
Transcript
MixTok. Adaptive Visual Tokenization with Global 1D Tokens and Sparse 2D Refinement. By Chenge Li and five others, from Zhejiang University, the Hong Kong University of Science and Technology, and one other institution. Published on the ninth of October. Its code repository is worked out: about one hundred and ten source files, with a readme.
A visual tokenizer has to balance two demands: keep the representation compact, and keep the fine-grained content of the image. The two common designs each lean one way. Grid-based tokenizers spend tokens uniformly across spatial locations, even though some regions of an image are far more complex than others. Compact one-dimensional tokenizers gather global information well, but can struggle to hold on to local detail.
MixTok is a hybrid that takes both. A compact one-dimensional global representation carries the image as a whole, and a sparse set of two-dimensional tokens is added for local refinement, only where it is needed.
Deciding where is the job of a content-adaptive router. It ranks candidate locations by refinement priority, and it sets how many two-dimensional tokens each image receives, under a target average token budget. A busy image can take more, a simple one fewer. At the highest-scoring locations, the two-dimensional token features replace the spatial features recovered from the one-dimensional representation. A single shared decoder then reconstructs the image from these mixed features.
On ImageNet at two hundred and fifty-six by two hundred and fifty-six pixels, MixTok is reported as highly competitive in both reconstruction and generation. Its reconstruction FID is one point one nine five, at an average of roughly one hundred and twenty-eight content tokens per image. Its generation FID is one point nine five, at an average of roughly one hundred and thirty-one content tokens.
The design question MixTok answers is where detail should live. A global summary handles what an image is broadly about, and a small, variable number of local tokens is spent only on the places that a uniform grid would have treated like everywhere else.
More with an account
Listen to every day’s papers
Daily playlists, deep dives, quests for the papers, and quizzes on their key terms.
Join free