GitHub - MCG-NJU/DDT: [CVPR 2026] DDT: Decoupled Diffusion Transformer

DDT: Decoupled Diffusion Transformer

Introduction

We decouple diffusion transformer into encoder-decoder design, and surprisingly that a more substantial encoder yields performance improvements as model size increases.

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

Checkpoints

We take the off-shelf VAE to encode image into latent space, and train the decoder with DDT.

Dataset	Model	Params	FID	HuggingFace
ImageNet256	DDT-XL/2(22en6de)	675M	1.26	🤗
ImageNet512	DDT-XL/2(22en6de)	675M	1.28	🤗

Online Demos

We provide online demos for DDT-XL/2(22en6de) on HuggingFace Spaces.

HF spaces: https://huggingface.co/spaces/MCG-NJU/DDT

To host the local gradio demo(default 512 resolution), run the following command:

# default 512 resolution
python app.py --config configs/repa_improved_ddt_xlen22de6_512.yaml --resolution 512 --ckpt_path=XXX512.ckpt
# for 256 resolution 
python app.py --config configs/repa_improved_ddt_xlen22de6_256.yaml --resolution 256 --ckpt_path=XXX256.ckpt

Usages

We use ADM evaluation suite to report FID.

# for installation
pip install -r requirements.txt

By default, the main.py will use all available GPUs. You can specify the GPU(s) to use with CUDA_VISIBLE_DEVICES. or specify the number of GPUs to use with as :

# in configs/repa_improved_ddt_xlen22de6_256.yaml
trainer:
  default_root_dir: universal_flow_workdirs
  accelerator: auto
  strategy: auto
  devices: auto
  # devices: 0,
  # devices: 0,1
  num_nodes: 1

By default, the save_image_callbacks will only save the first 100 images and npz file(to calculate FID with ADM suite). You can change the number of images to save with as :

# in configs/repa_improved_ddt_xlen22de6_256.yaml
callbacks:
- class_path: src.callbacks.model_checkpoint.CheckpointHook
  init_args:
    every_n_train_steps: 10000
    save_top_k: -1
    save_last: true
- class_path: src.callbacks.save_images.SaveImagesHook
  init_args:
     save_dir: val
     max_save_num: 0
     # max_save_num: 100

By default, we infer 50K images with batch size 64. You can change the number of images and classes to infer with as :

# in configs/repa_improved_ddt_xlen22de6_256.yaml
data:
  train_dataset: imagenet256
  train_root: /mnt/bn/wangshuai6/data/ImageNet/train
  train_image_size: 256
  train_batch_size: 16
  eval_max_num_instances: 50000
  pred_batch_size: 64
  # pred_batch_size: 16
  pred_num_workers: 4
  pred_seeds: null
  # pred_seeds: 1,2,3,4
  pred_selected_classes: null
  # pred_selected_classes:
  # - 0
  # - 1
  num_classes: 1000
  latent_shape:
  - 4
  - 32
  - 32

pred_selected_classes is a list of class indices to infer, if pred_selected_classes is null, all classes will be inferred. pred_seeds is string of seeds for every class seprated with , to infer

# for inference
python main.py predict -c configs/repa_improved_ddt_xlen22de6_256.yaml --ckpt_path=XXX.ckpt
# # or specify the GPU(s) to use with as :
CUDA_VISIBLE_DEVICES=0,1, python main.py predict -c configs/repa_improved_ddt_xlen22de6_256.yaml --ckpt_path=XXX.ckpt

# for training
# extract image latent (optional)
python3 tools/cache_imlatent4.py
# train
python main.py fit -c configs/repa_improved_ddt_xlen22de6_256.yaml

Reference

@article{wang2025ddt,
  title={DDT: Decoupled Diffusion Transformer},
  author={Wang, Shuai and Tian, Zhi and Huang, Weilin and Wang, Limin},
  journal={arXiv preprint arXiv:2504.05741},
  year={2025}
}

Acknowledgement

The code is mainly built upon FlowDCN, we also borrow ideas from the REPA, MAR and SiT.