Fast Sampling of Diffusion Models with Exponential Integrator

Authors: Qinsheng Zhang, Yongxin Chen

ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct comprehensive experiments to validate the efficacy of DEIS. For instance, with a pre-trained model (Song et al., 2020b), DEIS is able to reach 4.17 FID with 10 NFEs, and 2.86 FID with 20 NFEs on CIFAR10.
Researcher Affiliation Academia Qinsheng Zhang Georgia Institute of Technology qzhang419@gatech.edu Yongxin Chen Georgia Institute of Technology yongchen@gatech.edu
Pseudocode Yes Algorithm 1 t AB-DEIS
Open Source Code Yes Project page and code: https://qsh-zh.github.io/deis.
Open Datasets Yes For instance, with a pre-trained model (Song et al., 2020b), DEIS is able to reach 4.17 FID with 10 NFEs, and 2.86 FID with 20 NFEs on CIFAR10. (Abstract), 64 64 Celeb A (Liu et al., 2015) with pre-trained model from Song et al. (2020a), class-conditioned 64 64 Image Net (Deng et al., 2009) with pre-trained model (Dhariwal & Nichol, 2021)
Dataset Splits No No specific dataset split information (percentages, sample counts, or explicit methodology) was found.
Hardware Specification No No specific hardware details (e.g., GPU/CPU models, memory) were provided for the experimental setup.
Software Dependencies No We implemented our approach in Jax and Py Torch.
Experiment Setup Yes Due to numerical issues, we set ending time t0 in DMs during sampling a non-zero number. Song et al. (2020b) suggests t0 = 10-3 for VPSDE and t0 = 10-5 for VESDE. One such option is the quadratic timestep suggested in (Song et al., 2020a)