datasets:
- bayes-group-diffusion/GAS-teachers
license: mit
tags:
- arxiv:2510.17699
pipeline_tag: unconditional-image-generation
GAS: Improving Discretization of Diffusion ODEs via Generalized Adversarial Solver
This repository contains the implementation for GAS: Improving Discretization of Diffusion ODEs via Generalized Adversarial Solver, a method presented in the paper GAS: Improving Discretization of Diffusion ODEs via Generalized Adversarial Solver.
The work introduces a novel approach to accelerate sampling in diffusion models without compromising generation quality. The Generalized Solver (GS) offers a simpler parameterization of the ODE sampler, and when combined with adversarial training, forms the Generalized Adversarial Solver (GAS), which enhances detail fidelity and mitigates artifacts. This method aims to reduce the computational cost of diffusion model sampling from dozens to just a few function evaluations.
For detailed code, setup instructions, and examples, please refer to the official GitHub repository: https://github.com/3145tttt/GAS
How to use
To generate images from a trained GS checkpoint, you can use the generate.py script. Set the --checkpoint_path option to the path of your trained model checkpoint.
# Generate 50000 images using 2 GPUs and a checkpoint from checkpoint_path
torchrun --standalone --nproc_per_node=2 generate.py \
--config=configs/edm/cifar10.yaml \
--outdir=data/teachers/cifar10 \
--seeds=50000-99999 \
--batch=1024 \
--steps=4 \
--checkpoint_path=checkpoint_path
For a fair comparison and to avoid leakage of test seeds into the training dataset, we recommend using seeds 50000-99999 for all datasets except MS-COCO, which should use seeds 30000-59999.
Citation
@misc{oganov2025gasimprovingdiscretizationdiffusion,
title={GAS: Improving Discretization of Diffusion ODEs via Generalized Adversarial Solver},
author={Aleksandr Oganov and Ilya Bykov and Eva Neudachina and Mishan Aliev and Alexander Tolmachev and Alexander Sidorov and Aleksandr Zuev and Andrey Okhotin and Denis Rakitin and Aibek Alanov},
year={2025},
eprint={2510.17699},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2510.17699},
}