'Visual Anagrams' automatically generates various trompe l'oeil pictures using image generation AI
A research team at the University of Michigan has announced `` Visual Anagrams ,'' a technology that uses
[2311.17919] Visual Anagrams: Generating Multi-View Optical Illusions with Diffusion Models
https://arxiv.org/abs/2311.17919
Visual Anagrams
https://dangeng.github.io/visual_anagrams/
' Jigsaw Permutations ' is like a jigsaw puzzle, where you disassemble a picture and put it back together to create a different picture. For example, in the trompe l'oeil picture below, when you disassemble and reassemble the deer picture, you can see that it becomes a picture of a kitchen.
Einstein's portrait turned into a painting of a houseplant.
The following was created as an example of ' Flips and 180° Rotations ' which becomes a different picture when flipped 180 degrees. If you turn the old woman's face upside down, it becomes a picture of a woman in a dress.
When you flip a penguin picture 180 degrees, it becomes a giraffe face.
If you turn the picture of a snowy mountain 90 degrees clockwise, it becomes a picture of a horse. ' 90° Rotations ' creates a different picture when rotated 90 degrees.
A picture of a mountain hut becomes a picture of a sailing ship.
`` Color Inversions '' is a trompe l'oeil painting in which black and white are reversed to create a different picture. For example, in the example below, the left side looks like a rabbit photo, and the right side looks like a teddy bear photo.
On the left is the profile of a sad woman, and on the right is the profile of a slightly smiling man.
Miscellaneous Permutations is a type of trompe l'oeil in which when a part of a picture is transformed or rotated, it becomes a different picture. For example, if you rotate just the face of a portrait of Marilyn Monroe, it becomes a portrait of Einstein.
I thought it was a picture of a leather chair, but when I rotated the center, it turned out to be a portrait of a man with a beard.
`` Random Patch Permutations '' is a technique in which a picture is broken down into small pieces and reassembled to create a different picture. For example, if you take the rabbit mosaic picture below apart and reassemble it, it becomes a duck picture.
When a painting depicting a young man was disassembled and reconstructed, it transformed into a painting of an old man.
The figure below shows the process of Visual Anagrams. Using a general diffusion model, noise is generated for each trompe l'oeil pattern, and the noise estimates are averaged.
The source code of Visual Anagrams is published in the GitHub repository below.
GitHub - dangeng/visual_anagrams: Code for the paper 'Visual Anagrams: Generating Multi-View Optical Illusions with Diffusion Models'
https://github.com/dangeng/visual_anagrams
Related Posts: