Mingkai Jia^1,2, Mingxiao Li², Zhijian Shu^2,3, Anlin Zheng⁴, Liaoyuan Fan², Jiaxin Guo⁵, Tianxing Shi³, Dongyue Lu², Zeming Li¹, Xiaoyang Guo², Xiaojuan Qi⁴, Xiao-Xiao Long³, Qian Zhang², Ping Tan^1*, Wei Yin^2*§,

HKUST¹, Horizon Robotics², NJU³, HKU⁴, CUHK⁵,
^* Corresponding Author, ^§ Project Leader

• [April 2026] Released Inference Code
• [April 2026] Released models & stats.
• [Nov 2025] Released paper.

• Training code.
• Models & Evaluation code.
• Huggingface models & stats.

git clone https://github.com/MKJia/DINO-Tok.git
cd DINO-Tok

conda create -n dinotok python=3.10
conda activate dinotok
pip3 install -r requirements.txt

Download the pretrained models & stats from our model & stat to your /path/to/your/ckpt.

We default use the ImageNet-1k dataset. Or you can try our UHDBench dataset on huggingface and download to your /path/to/your/dataset.

Remember to change the paths in scripts.

bash scripts/test_aetok.bash
bash scripts/test_aegen.bash
bash scripts/test_vqtok.bash
bash scripts/test_vqgen.bash

• 🔥 Qualitative reconstruction images.

• 🔥 Qualitative class-to-image generation of Imagenet.

• 🔥 Evaluation of dino-tok-ae on 256×256 ImageNet benchmark.

• 🔥 Evaluation of dino-tok-vq on 256×256 ImageNet benchmark.

If the paper and code from DINO-Tok help your research, we kindly ask you to give a citation to our paper ❤️. Additionally, if you appreciate our work and find this repository useful, giving it a star ⭐️ would be a wonderful way to support our work. Thank you very much.

@article{jia2025dinotok,
  title={DINO-Tok: Adapting DINO for Visual Tokenizers},
  author={Jia, Mingkai and Li, Mingxiao and Fan, Liaoyuan and Shi, Tianxing and Guo, Jiaxin and Li, Zeming and Guo, Xiaoyang and Long, Xiao-Xiao and Zhang, Qian and Tan, Ping and others},
  journal={arXiv preprint arXiv:2511.20565},
  year={2025}
}

This repository is under the MIT License. For more license questions, please contact Mingkai Jia (mjiaab@connect.ust.hk) and Wei Yin (yvanwy@outlook.com).