MixTok: Adaptive Visual Tokenization
with Global 1D Tokens and Sparse 2D Refinement

Chenge Li¹Yefei He^1,†,‡Xinyu Liu²Feng Chen^3,†,‡Hong Zhou¹Bohan Zhuang¹

¹Zhejiang University ²The Hong Kong University of Science and Technology ³University of Adelaide

† Project lead‡ Corresponding author

billhe@zju.edu.cn chenfeng1271@gmail.com

Paper · arXiv Code Models

MixTok class-conditional generation examples showing base images, selected refinement locations, and final refined images — **Content-adaptive generation.** Each group shows the 1D base image, selected refinement locations, and the final image after sparse 2D refinement.

1.195

rFID at ≈128 tokens

1.9511

gFID at ≈131 tokens

270.1 ms

RTX 5090 end-to-end latency

Abstract

Efficient visual tokenization requires compact representations without losing fine-grained image content. Dense grid tokenizers allocate capacity uniformly, while short 1D token sequences capture global structure but can miss local detail. MixTok combines both: a compact 1D representation is expanded into a base feature grid, and a content-adaptive router selects how many 2D tokens to retain and where to place them. Selected 2D features replace the corresponding base features before a shared decoder reconstructs the image.

The same representation supports two-stage masked generation. A 1D MaskGIT first generates global tokens; a distilled grid scorer chooses the sparse refinement set; a 2D MaskGIT then generates local tokens using continuous spatial context from the base representation. On ImageNet 256×256, MixTok reaches an rFID of 1.195 and a gFID of 1.9511 with average budgets of roughly 128 and 131 content tokens.

Method

Reconstruction overview

MixTok reconstruction architecture with 1D and 2D token branches, reconstruction grid scorer, budget controller, and shared decoder — A compact 1D representation provides the base feature grid, while the router selects sparse 2D features for local refinement. The shared decoder reconstructs the image from the mixed feature grid.

Global base. A short 1D token sequence is spatially expanded and aligned with the shared decoder’s feature space.
Adaptive routing. A grid scorer ranks candidate positions; top-p selection jointly determines the locations and per-image token count.
Local refinement. Selected 2D token features replace the aligned base features, and the shared decoder renders the mixed grid.

Two-stage masked generation

Two-stage MixTok generation with a 1D MaskGIT, generation grid scorer, embedding adapter, sparse 2D MaskGIT, codebook lookup projection, and feature mixing — During generation, the 2D branch is unavailable to the router. A Generation Grid Scorer distilled from its reconstruction counterpart predicts the refinement set from the base grid alone. An Embedding Adapter maps base features into spatial context embeddings for the sparse 2D MaskGIT.

Results

MixTok is evaluated on ImageNet 256×256. Reconstruction uses all 50,000 validation images; MixTok’s generation gFID is averaged over four independent 50,000-image sampling runs.

Image reconstruction

Tokenizer	Avg. tokens	rFID ↓	PSNR ↑
TiTok S-128	128	1.7222	17.72
TiTok BL-128 VQ	128	1.5063	19.83
One-D-Piece L-256	128	1.4406	17.71
FlexTok d18-d28	128	1.1422	16.76
DetailFlow-16	128	1.1370	19.46
MixTok	128	1.1950	19.51

Class-conditional generation and efficiency

Generator	Params	Tokens	Steps	gFID ↓	Latency ↓	FLOPs ↓
AdaGen-MaskGIT-L	206M	256	16	2.4883	372.0 ms	3.623 T
TiTok-S-128	295M	128	64	1.9994	1018.4 ms	9.949 T
Halton-MaskGIT-L†	480M	576	32	2.5143	591.5 ms	26.690 T
One-D-Piece-L-256	177M	256	16	2.4144	350.1 ms	3.526 T
LPD-L-256	337M	256	20	2.3863	215.5 ms	0.879 T
LlamaGen-L	343M	256	256	3.6803	4951.8 ms	0.599 T
DetailFlow-32	326M	256	39	2.6315	669.9 ms	5.627 T
MixTok	321M	130.66	8+8	1.9511	270.1 ms	1.452 T

Latency is median end-to-end generation time at batch size 1 on one RTX 5090, including RGB decoding. FLOPs are counted per image with multiply-add counted as two. †Halton-MaskGIT-L uses its native 384×384 output; other rows use 256×256.

3.77× faster than TiTok-S-128 under the measured single-image protocol. 2.48× faster than DetailFlow-32 while producing fewer content tokens. 8+8 masked sampling rounds in the main generation configuration.

Visualizations

Reconstruction

Image-dependent token budgets

Distribution of selected 2D token counts for reconstruction and three generation seeds — MixTok varies the refinement budget by image. Generation distributions remain stable across sampling seeds.

BibTeX

If you find MixTok useful, please cite our paper.

@misc{li2026mixtok,
  title  = {MixTok: Adaptive Visual Tokenization with Global 1D Tokens
            and Sparse 2D Refinement},
  author = {Li, Chenge and He, Yefei and Liu, Xinyu and Chen, Feng
            and Zhou, Hong and Zhuang, Bohan},
  year   = {2026},
  eprint = {2610.13181},
  archivePrefix = {arXiv},
  primaryClass = {cs.CV},
  url    = {https://arxiv.org/abs/2610.13181}
}

MixTok: Adaptive Visual Tokenization with Global 1D Tokens and Sparse 2D Refinement