Diffusable Latents from Structure-Agnostic Distillation
The paper explores distilling large foundation models into autoencoder latents using structure-agnostic methods instead of position-wise matching. It proposes pooling image latent tokens into a single descriptor, which improves latent diffusability for faster and higher quality diffusion model sampling. Results show this approach performs comparably to or better than dense position-wise distillation across various latent shapes and teacher modalities.