DiC: Rethinking Conv3x3 Designs in Diffusion Models

June 12, 2025 · View on GitHub

ConvUNets have been overlooked... but they outperform Diffusion Transformers!

News

6/11/2025: We have released the codes of DiC! 🔥🔥🔥 Weights, SiT, and REPA versions are coming very soon.

3/3/2025: Codes & Weights are at the final stage of inspection. We will have them released ASAP.

2/27/2025: DiC is accepted by CVPR 2025! 🎉🎉

effect

🤔 In this work, we intend to build a diffusion model with Conv3x3 that is simple but efficient.

🔧 We re-design architectures & blocks of the model to tap the potential of Conv3x3 to the full.

🚀 The proposed DiC ConvUNets are more powerful than DiTs, and much much faster!

Repo Outline

This repo is mostly based on the official repo of DiT. Weights, SiT and REPA versions will be opensourced very soon.

Torch model script: dic_models.py

Preparation

Please run command pip install -r requirements.txt to install the supporting packages.

(Optional) Please download the VAE from this link. The VAE could be automatically downloaded as well.

Training

Here we provide two ways to train a DiC model: 1. train on the original ImageNet dataset; 2. train on preprocessed VAE features (Recommended).

Training Data Preparation Use the original ImageNet dataset + VAE encoder. Firstly, download ImageNet as follows:

imagenet/
├──train/
│  ├── n01440764
│  │   ├── n01440764_10026.JPEG
│  │   ├── n01440764_10027.JPEG
│  │   ├── ......
│  ├── ......
├──val/
│  ├── n01440764
│  │   ├── ILSVRC2012_val_00000293.JPEG
│  │   ├── ILSVRC2012_val_00002138.JPEG
│  │   ├── ......
│  ├── ......

Then run the following command:

torchrun --nnodes=1 --nproc_per_node=8 train.py --data-path={path to imagenet/train} --image-size=256 --model={model name} --epochs={iteration//5000} # fp32 Training

accelerate launch --mixed_precision fp16 train_accelerate.py --data-path {path to imagenet/train} --image-size=256 --model={model name} --epochs={iteration//5000} # fp16 Training

Training Feature Preparation (RECOMMENDED)

Following Fast-DiT, it is recommended to load VAE features directly for faster training. You don't need to download the enormous ImageNet dataset (> 100G); instead, a much smaller "VAE feature" dataset (~21G for ImageNet 256x256) is available here on HuggingFace and MindScope. Please do the following steps:

Download imagenet_feature.tar
Unzip the tar ball by running tar -xf imagenet_feature.tar

imagenet_feature/
├── imagenet256_features/ # VAE features
└── imagenet256_labels/ # labels

Append parser --feature-path={path to imagenet_feature} to the training command.

torch --nnodes=1 --nproc_per_node=8 sample_ddp.py --ckpt={path to checkpoint} --image-size=256 --model={model name} --cfg-scale={cfg scale}

BibTex Formatted Citation

If you find this repo useful, please cite:

@article{tian2025dic,
  author       = {Yuchuan Tian and
                  Jing Han and
                  Chengcheng Wang and
                  Yuchen Liang and
                  Chao Xu and
                  Hanting Chen},
  title        = {DiC: Rethinking Conv3x3 Designs in Diffusion Models},
  journal      = {CoRR},
  volume       = {abs/2501.00603},
  year         = {2025},
  url          = {https://doi.org/10.48550/arXiv.2501.00603},
  doi          = {10.48550/ARXIV.2501.00603},
  eprinttype    = {arXiv},
  eprint       = {2501.00603},
  timestamp    = {Mon, 10 Feb 2025 21:52:20 +0100},
  biburl       = {https://dblp.org/rec/journals/corr/abs-2501-00603.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}