---
license: apache-2.0
datasets:
- gaunernst/ffhq-1024-wds
---
# MADFormer-FFHQ

This repository provides checkpoints for MADFormer trained on **FFHQ-1024**, combining autoregressive global conditioning and diffusion-based local refinement for high-resolution image synthesis.

---

## 📄 Paper

[MADFormer: Mixed Autoregressive & Diffusion Transformers for Continuous Image Generation](https://arxiv.org/abs/2506.07999)

---

## 📦 Checkpoints

- Trained for **210k steps** on FFHQ-1024  
- Download checkpoint: `ckpts.pt`

---

## 🧪 How to Use

```python
# TODO
```

> 💡 MADFormer supports flexible AR↔Diff trade-offs. On FFHQ-1024, increasing AR layer allocation yields up to **75% FID improvements** under low NFE settings.

---

## 📚 Citation

If you find our work useful, please cite:

```bibtex
@misc{chen2025madformermixedautoregressivediffusion,
      title={MADFormer: Mixed Autoregressive and Diffusion Transformers for Continuous Image Generation}, 
      author={Junhao Chen and Yulia Tsvetkov and Xiaochuang Han},
      year={2025},
      eprint={2506.07999},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2506.07999}, 
}
```