or: Subspace Diffusion Generative Models

Use a modified PCA to determine which components of the diffusing vector are gaussian, and which are more interesting. Only bother learning the diffusion process in the interesting subspaces. Discarding half the latent space seems possible with no significant loss of image quality.