arXiv 2502.14831

Improving the Diffusability of Autoencoders

By Ivan Skorokhodov, Sharath Girish, et al.

Published 2025-02-20

Discussion

Read the public discussion and references gathered around this paper.

Latent diffusion models have emerged as the leading approach for generating high-quality images and videos, utilizing compressed latent representations to reduce the computational burden of the diffusion process. While recent advancements have primarily focused on scaling diffusion backbones and improving autoencoder reconstruction quality, the interaction between these components has received comparatively less att…

View the original paper on arXiv