Unlike prior autoencoder-based diffusion models ... most of which are in the convolutional layers. Unlike the text encoder, the VAE is activation-heavy. Using the same 8 MB cut, an encoder with a ...