metadata
license: mit
pipeline_tag: image-to-image
tags:
- discrete tokenization
- autoregressive generation
InsightTok
InsightTok is a discrete visual tokenizer designed to improve the fidelity of text and faces, two of the most challenging yet perceptually important structures in autoregressive image generation.
It was introduced in the paper InsightTok: Improving Text and Face Fidelity in Discrete Tokenization for Autoregressive Image Generation.
- Paper: InsightTok: Improving Text and Face Fidelity in Discrete Tokenization for Autoregressive Image Generation
- Code: https://github.com/LeapLabTHU/InsightTok
Model Details
| Property | Value |
|---|---|
| Downsampling rate | 16× |
| Codebook size | 16,384 |
| Latent dimension | 256 |
| Number of parameters | 426M |
Performance
InsightTok achieves strong text and face reconstruction quality while maintaining a compact discrete representation through localized, content-aware perceptual losses.
Usage
InsightTok follows the standard VQGAN-style autoencoding interface. For setup and implementation details, please refer to the GitHub repository.
# image encoding
latents, _, [_, _, indices] = vq_model.encode(input_image_tensor)
# image decoding
recon_image_tensor = vq_model.decode(latents)
Citation
@article{yue2026insighttok,
title={InsightTok: Improving Text and Face Fidelity in Discrete Tokenization for Autoregressive Image Generation},
author={Yue, Yang and Wei, Fangyun and He, Tianyu and Zhao, Jinjing and Ni, Zanlin and Liu, Zeyu and Guo, Jiayi and Shi, Lei and Dong, Yue bit and Chen, Li and Li, Ji and Huang, Gao and Chen, Dong},
journal={arXiv preprint arXiv:2605.14333},
year={2026}
}