lossfunk Research: Diffusion Models Internally Represent Visual Illusions

paraschopra · x · 2026-08-03

New research from @lossfunk reveals that diffusion (denoising) models internally represent visual illusions.

The study tested 9 denoising models across three major architecture families: pixel-space DDPMs, latent diffusion U-Nets, and diffusion transformers.

The findings show that all evaluated models developed internal representations capable of discriminating visual illusions from physically identical control regions.

Original post →

More from Research

Research channel →