Glossary

Noise

UNet

  • GroupNorm applies group normalization to the inputs of each block
  • Dropout layers for smoother training
  • Multiple resnet layers per block (if layers_per_block isn't set to 1)
  • Attention (usually used only at lower resolution blocks)
  • Conditioning on the timestep.
  • Downsampling and upsampling blocks with learnable parameters

DDPM

Schedulers or Samplers

Sources