Beyond Heat Dissipation: Optimizing Diffusion Models in Frequency Domain

Code for Beyond Heat Dissipation: Optimizing Diffusion Models in Frequency Domain (TPAMI-2026).

Installation

conda env create -f environment.yml
conda activate fibbo
cd fibbo

Note: For the DDPM++ architecture, more extentional packages should be installed, which is listed in requirements_ddpmpp.txt.

Data

The datasets will be automatically downloaded to ./data/cifar10, ./data/cifar100, ./data/mnist during running.

Running

We provide training and evaluation scripts in ./scripts. The project supports three network architectures across multiple datasets:

DDPM Architecture

Automatically evaluating the metric of the generation after training.

CIFAR-10

bash scripts/train_ddpm_cifar10.sh

CIFAR-100

bash scripts/train_ddpm_cifar100.sh

MNIST

bash scripts/train_ddpm_mnist.sh

BDM Architecture

CIFAR-10

bash scripts/train_bdm_cifar10.sh
bash scripts/eval_bdm_cifar10.sh

DDPM++ Architecture

CIFAR-10

bash scripts/train_ddpmpp_cifar10.sh
bash scripts/eval_ddpmpp_cifar10.sh

CIFAR-100

bash scripts/train_ddpmpp_cifar100.sh
bash scripts/eval_ddpmpp_cifar100.sh

Acknowledgement

We appreciate the repositories pytorch-ddpm, Score-based SDE for references when implementing this project. We also thank other projects for any assistance they may have provided. Thanks to all the authors for their great contributions.

Citation

@article{wang2026beyond,
  title={Beyond Heat Dissipation: Optimizing Diffusion Models in Frequency Domain},
  author={Wang, Qisen and Zhao, Yifan and Li, Jia},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year={2026},
  publisher={IEEE}
}