Rethinking Timesteps Samplers and Prediction Types

Bin Xie, Gady Agam

Diffusion models suffer from the huge consumption of time and resources to train. For example, diffusion models need hundreds of GPUs to train for several weeks for a high-resoluti…