Presenter

Details

  • Date: Monday, April 15, 2024
  • Time: 12:00 PM
  • Location: EECS, room 2311

Abstract

Recent work has suggested that state-of-the-art computer vision models can be “tricked” by optical illusions just as humans are. In this talk, I take this observation a step further and show that these models can actually synthesize optical illusions as well, with no finetuning. We take an off-the-shelf diffusion model and adapt the sampling procedure to generate what we call “visual anagrams”: analogues of textual anagrams in which an image changes appearance upon a permutation of its pixels. These include images that change appearance when rotated, flipped, or color inverted, or jigsaw puzzles that can be solved in two different ways.

Registration

Please fill out this survey to RSVP for pizza!

For any additional questions or inquiries, please contact us at speecs.seminar-requests@umich.edu