# DDT: Decoupled Diffusion Transformer
[![arXiv](https://img.shields.io/badge/arXiv-2504.05741-b31b1b.svg)](https://arxiv.org/abs/2504.05741)
[![Paper page](https://huggingface.co/datasets/huggingface/badges/resolve/main/paper-page-sm.svg)](https://huggingface.co/papers/2504.05741)

[![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/ddt-decoupled-diffusion-transformer/image-generation-on-imagenet-256x256)](https://paperswithcode.com/sota/image-generation-on-imagenet-256x256?p=ddt-decoupled-diffusion-transformer)

[![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/ddt-decoupled-diffusion-transformer/image-generation-on-imagenet-512x512)](https://paperswithcode.com/sota/image-generation-on-imagenet-512x512?p=ddt-decoupled-diffusion-transformer)
## Introduction
We decouple diffusion transformer into encoder-decoder design, and surpresingly that a **more substantial encoder yields performance improvements as model size increases**.
![](./figs/main.png)
* We achieves **1.26 FID** on ImageNet256x256 Benchmark with DDT-XL/2(22en6de).
* We achieves **1.28 FID** on ImageNet512x512 Benchmark with DDT-XL/2(22en6de).
* As a byproduct, our DDT can reuse encoder among adjacent steps to accelerate inference.
## Visualizations
![](./figs/teaser.png)
## Checkpoints
Waiting for release.

## Usgae
We use ADM evaluation suite to report FID.
```bash
# for installation
pip install -r requirements.txt
```

```bash
# for training
python main.py fit -c configs/repa_improved_ddt_xlen22de6_256.yaml
```

```bash
# for inference
python main.py predict -c configs/repa_improved_ddt_xlen22de6_256.yaml --ckpt_path=XXX.ckpt
```
## Reference
```bibtex
@ARTICLE{ddt,
  title         = "DDT: Decoupled Diffusion Transformer",
  author        = "Wang, Shuai and Tian, Zhi and Huang, Weilin and Wang, Limin",
  month         =  apr,
  year          =  2025,
  archivePrefix = "arXiv",
  primaryClass  = "cs.CV",
  eprint        = "2504.05741"
}
```