Deep Compression Autoencoder From paper "Deep Compression Autoencoder for Efficient High-Resolution Diffusion Models" There are three version: Three phase training with AdamW: [Link] One phase training with AdamW: [Link] One phase training with SOAP: [Link]