README.md

---
pipeline_tag: unconditional-image-generation
---

# Image Tokenizer Needs Post-Training

This repository contains **RobusTok**, a novel image tokenizer presented in the paper [Image Tokenizer Needs Post-Training](https://huggingface.co/papers/2509.12474).

<div align="center">

[![Project Page](https://img.shields.io/badge/%20project%20page-lightblue)](https://qiuk2.github.io/works/RobusTok/index.html)&nbsp;
[![GitHub Code](https://img.shields.io/badge/%20GitHub%20Code-brightgreen)](https://github.com/qiuk2/RobusTok)&nbsp;
[![🤗 Weights](https://img.shields.io/badge/%F0%9F%A4%97%20Weights-yellow)](https://huggingface.co/qiuk6/RobusTok)&nbsp;

</div>

<div align="center">
  <img src="https://github.com/qiuk2/RobusTok/raw/main/assets/teaser.png" alt="Teaser" width="95%">
</div>

---

## About RobusTok

Recent image generative models typically rely on a frozen image tokenizer to capture the image distribution in a latent space. However, a significant discrepancy exists between the reconstruction and generation distribution, as current tokenizers often prioritize the reconstruction task without fully considering generation errors during sampling.

**RobusTok** addresses this by proposing a novel tokenizer training scheme that includes both main-training and post-training:
*   **Main training:** Constructs a robust latent space by simulating sampling noises and unexpected tokens.
*   **Post-training:** Further optimizes the tokenizer decoder with respect to a well-trained generative model, mitigating the distribution difference between generated and reconstructed tokens.

This approach significantly enhances the robustness of the tokenizer, boosting generation quality and convergence speed.

## Key Highlights of Post-Training

- 🚀 **Better generative quality**: Achieves notable improvements in gFID (e.g., 1.60 gFID → 1.36 gFID with a ~400M generator).
- 🔑 **Generalizability**: Applicable to both autoregressive & diffusion models.
- ⚡ **Efficiency**: Provides strong results with relatively small generative models.

## Model Zoo

| Generator \ Tokenizer | RobusTok w/o. P.T | RobusTok w/. P.T |
|---|---:|---:|
| Base ([weights](https://huggingface.co/qiuk6/RobusTok/resolve/main/rar_b.bin?download=true)) | gFID = 1.83 | gFID = 1.60 |
| Large ([weights](https://huggingface.co/qiuk6/RobusTok/resolve/main/rar_l.bin?download=true)) | gFID = 1.60 | gFID = 1.36 |

## Usage

Due to the specialized nature of RobusTok's tokenizer and generator training and inference pipeline, detailed usage instructions, installation guides, and code examples are provided in the [official GitHub repository](https://github.com/qiuk2/RobusTok). This includes scripts for:

*   Environment setup and package installation.
*   Dataset preparation.
*   Main training for the tokenizer.
*   Training code for the generator.
*   Post-training for the tokenizer.
*   Inference and evaluation (see [Inference Code](https://github.com/qiuk2/RobusTok#inference-code)).

## Visualization

<div align="center">
  <img src="https://github.com/qiuk2/RobusTok/raw/main/assets/ft-diff.png" alt="vis" width="95%">
  <p>
    Visualization of 256&times;256 image generation before (top) and after (bottom) post-training. Three improvements are observed: (a) OOD mitigation, (b) Color fidelity, (c) detail refinement.
  </p>
</div>

---

## Citation

If our work assists your research, feel free to give us a star ⭐ or cite us using:

```bibtex
@misc{qiu2025imagetokenizerneedsposttraining,
      title={Image Tokenizer Needs Post-Training},
      author={Kai Qiu and Xiang Li and Hao Chen and Jason Kuen and Xiaohao Xu and Jiuxiang Gu and Yinyi Luo and Bhiksha Raj and Zhe Lin and Marios Savvides},
      year={2025},
      eprint={2509.12474},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2509.12474},
}
```