This article discusses a method for generating multi-view optical illusions using pre-trained diffusion models. These illusions are images that appear as one thing but change appearance or identity when transformed. The method supports a wide range of transformations such as rotations, flips, color inversions, jigsaw rearrangements, and random permutations. The authors explain their approach, which involves estimating noise in different views or transformations of an image and aligning and averaging the noise estimates. They also discuss the conditions and constraints that need to be met for this method to work effectively. Overall, this research provides a simple yet effective way to create visually captivating optical illusions.
https://dangeng.github.io/visual_anagrams/