ECCV 2026 paper: x0-prediction tackles high-dimensional latent diffusion

An ECCV 2026 paper studies the diffusability of representation autoencoder (RAE) latent spaces, showing that fine-tuning encoders for reconstruction lowers effective dimensionality, and that x0-prediction outperforms velocity prediction for high-dimensional latent diffusion, improving text-to-image generation.

2026-09-24 ~ 2026-09-24 · 2 related posts