edm2
The official PyTorch code for the CVPR 2024 paper presents improvements in training dynamics of diffusion models for image synthesis. By addressing inefficiencies in the ADM diffusion model, the paper suggests network redesigns to maintain activation and weight balance without changing the overall structure. These optimizations improve FID scores from 2.41 to 1.81 on ImageNet-512, using deterministic sampling. A new method for post-training EMA parameter tuning is also introduced, enabling precise adjustments without extra training runs.
