SEEDS: Exponential SDE Solvers for Fast High-Quality Sampling from Diffusion Models

Official PyTorch implementation of the SEEDS paper (NeurIPS 2023).

Martin Gonzalez, Nelson Fernandez, Thuy Tran, Elies Gherbi, Hatem Hajri and Nader Masmoudi.

Image sampled from Stable Diffusion XL with SEEDS at 200 NFE.

Overview

We address the challenges of slow SDE sampling in Diffusion Probabilistic Models (DPMs) by introducing Stochastic Exponential Derivative-free Solvers (SEEDS). SEEDS are designed to provide optimal quality sampling without the need for a large number of evaluations (NFEs). Previous efforts have focused on improving speed by crafting powerful ODE solvers, but they often fall short of achieving the optimal quality obtained by slower SDE solvers. We accomplish this by analytically computing the linear part of solutions in diffusion SDEs and incorporating innovative techniques for handling stochastic components. Inspired by the Stochastic Exponential Time-Differencing method, SEEDS include Markov preserving high-order terms that significantly accelerate the sampling process. Importantly, SEEDS do not require derivatives or training. This research provides the first set of scalable and efficient solvers with fully proven strong & weak convergence guarantees. We validate our approach on multiple image generation benchmarks, demonstrating that SEEDS is about 3-5 times faster than previous SDE methods while either outperforming or remaining competitive with their quality.

Requirements

We recommend running on Linux for performance and compatibility reasons.
1+ GPU should be used for sampling. We have done all testing and development using Tesla V100S GPUs.
64-bit Python 3.9 and PyTorch 1.13.0 (or later) which includes the function torch.torch.distributed.all_gather_into_tensor() required to compute the Inception score. See https://pytorch.org for PyTorch install instructions.
Python libraries: The required libraries are described in the file requirements.txt. You can easily install using the following pip command:

pip install -r requirements

More details coming soon

Details on how to run our code will be soon available.

License

Our code is built on top of the EDM code Elucidating the Design Space of Diffusion-Based Generative Models (EDM) under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

Citation

If you find our work useful, please consider citing:

@article{gonzalez2023seeds,
  author    = {Gonzalez, Martin and Fernandez Pinto, Nelson and Tran, Thuy and Gherbi, Elies and Hajri, Hatem and Masmoudi, Nader},
  title     = {SEEDS: Exponential SDE Solvers for Fast High-Quality Sampling from Diffusion Models},
  journal   = {arXiv preprint arXiv:2305.14267},
  year      = {2023}
}

nfsrules / SEEDS