---
pipeline_tag: image-to-image
---

# NativeTok: Native Visual Tokenization for Improved Image Generation

This repository contains the official weights for NativeTok, a framework that enforces causal dependencies during tokenization to improve generative modeling coherence and performance.

[**Paper**](https://huggingface.co/papers/2601.22837) | [**GitHub**](https://github.com/wangbei1/Nativetok)

## Introduction

NativeTok consists of a Meta Image Transformer (MIT) for latent image modeling and a Mixture of Causal Expert Transformer (MoCET), where each lightweight expert block generates a single token conditioned on prior tokens and latent features. This approach addresses the mismatch between tokenization and generative modeling by embedding relational constraints within token sequences.

## 📌 Model Checkpoints

These weights are trained based on the MaskGIT architecture with OrderTok tokenization strategies.

| File Name | Description | Resolution |
| :--- | :--- | :--- |
| **`maskgit128_ordertok.bin`** | MaskGIT | 256x256 |
| **`Nativetok_128_300000_stage2.bin`** | Nativetok_128 checkpoint | 256x256 |

## 🚀 Quick Start

### Download Weights
You can use `huggingface_hub` to download the weights directly:

```python
from huggingface_hub import hf_hub_download

# Download the main weight
checkpoint_path = hf_hub_download(
    repo_id="wangbei1/Nativetok", 
    filename="maskgit128_ordertok.bin"
)

print(f"Model downloaded to: {checkpoint_path}")
```

## Related Resources
Base Framework (1D-Tokenizer): [1D-Tokenizer](https://yucornetto.github.io/projects/titok.html)

## Citation

```bibtex
@article{wu2026nativetok,
  title={NativeTok: Native Visual Tokenization for Improved Image Generation},
  author={Bin Wu and Mengqi Huang and Weinan Jia and Zhendong Mao},
  journal={arXiv preprint arXiv:2601.22837},
  year={2026}
}
```