This repository contains the official implementation of the paper:
Learning to Integrate Diffusion ODEs by Averaging the Derivatives.
This work introduces an intermediate strategy for accelerating diffusion model inference by learning ODE integration. It uses loss functions derived from the derivative-integral relationship, inspired by Monte Carlo integration and Picard iteration. Referred to as "secant losses," this approach balances performance and cost, leading to great training stability. The secant version of EDM achieves a 10-step FID of 2.14 on CIFAR-10, while the secant version of SiT-XL/2 attains a 4-step FID of 2.27 and an 8-step FID of 1.96 on ImageNet-256x256.
For installation instructions and how to train and evaluate the models, please refer to the
GitHub repository.
Trained checkpoints for different settings (SDEI, STEE, SDEE, STEE, various step counts) are provided on
Hugging Face.
1@article{liu2025learning,
2 title={Learning to Integrate Diffusion ODEs by Averaging the Derivatives},
3 author={Liu, Wenze and Yue, Xiangyu},
4 journal={arXiv preprint arXiv:2505.14502},
5 year={2025}
6}
