The Art of Deception: Color Visual Illusions and Diffusion Models

TL;DR

This paper investigates how diffusion models encode and generate visual illusions, revealing human-like perception shifts and demonstrating their ability to produce new illusions that can deceive both models and humans.

Contribution

It is the first to show that diffusion models encode visual illusions similarly to humans and can generate new illusions in realistic images using text-to-image models.

Findings

Diffusion models exhibit human-like brightness and color shifts in their latent space.

They can predict and generate visual illusions in images.

Model-generated illusions can also fool human observers.

Abstract

Visual illusions in humans arise when interpreting out-of-distribution stimuli: if the observer is adapted to certain statistics, perception of outliers deviates from reality. Recent studies have shown that artificial neural networks (ANNs) can also be deceived by visual illusions. This revelation raises profound questions about the nature of visual information. Why are two independent systems, both human brains and ANNs, susceptible to the same illusions? Should any ANN be capable of perceiving visual illusions? Are these perceptions a feature or a flaw? In this work, we study how visual illusions are encoded in diffusion models. Remarkably, we show that they present human-like brightness/color shifts in their latent space. We use this fact to demonstrate that diffusion models can predict visual illusions. Furthermore, we also show how to generate new unseen visual illusions in realistic images using text-to-image diffusion models. We validate this ability through psychophysical experiments that show how our model-generated illusions also fool humans.