prior-aligned-autoencoders

This article introduces Prior-Aligned Autoencoders (PAE), a new method for creating diffusion-friendly latent manifolds that achieves state-of-the-art image generation quality while enabling 13x faster training convergence.