ECCV 2026 paper studies which high-dimensional latents suit diffusion models

_akhaliq · x · 2026-09-25

An ECCV 2026 paper from Adobe researchers and Andrew Owens examines Representation Autoencoders (RAEs) that let diffusion models operate in pretrained visual encoder feature spaces — many off-the-shelf encoders discard fine-grained detail, and the paper systematically studies which high-dimensional latents are diffusible.

Related event: Adobe ECCV paper tames diffusion in high-dimensional latent spaces(3 posts)→

Original post →

More from Research

Research channel →