/mini_edm

Minimum implementation of EDM (Elucidating the Design Space of Diffusion-Based Generative Models) on cifar10 and mnist

Primary LanguagePython

Mini-EDM -- A minimalistic EDM with all the essence

Introduction

To train or evaluate the EDM, please follow the command in run.sh.

Results

Results FID on CIFAIR-10:

FID of this implementation.

Reference FID (each snapshot=5000k images=5000k/128 steps) from the original EDM codebase using the same model architecture and the same batch size of 128 (nothing else changed):

FID of original EDM.

The difference may come from the learning rate and warmup strategy.