FLAIR · project
DEIS: Fast Sampling of Diffusion Models with Exponential Integrator
Diffusion models can generate high-quality samples, but repeatedly evaluating a large neural network during denoising makes inference costly. DEIS asks how much of this cost can be reduced by improving the numerical solver while reusing a pretrained model.
Exploiting the diffusion dynamics
The sampling dynamics have a semilinear structure: a known linear component is combined with a learned nonlinear component. DEIS uses exponential integration to handle that structure directly, reducing the discretization error that arises when the continuous process is approximated with a small number of steps.
Faster sampling from pretrained models
The paper evaluates the trade-off between sample quality and the number of neural-network evaluations, demonstrating image generation with substantially fewer sampling steps. Its contribution is an inference algorithm that works with pretrained diffusion models without requiring them to be retrained. The code includes the sampler and configurations needed to examine the quality–computation trade-off.
Publication
Fast Sampling of Diffusion Models with Exponential Integrator
Qinsheng Zhang and Yongxin Chen
ICLR 2023