Visual Anagrams: Generating optical illusions with diffusion models (dangeng.github.io)

Article summary

The article presents a method for generating multi-view optical illusions using diffusion models. This method can create images that change appearance when transformed in various ways, such as rotations, flips, and color inversions. The technique uses a pre-trained diffusion model to estimate noise in different views of an image and then combines these estimates to generate the illusion. The authors demonstrate the method's capabilities with various examples, including jigsaw puzzles and multiple-view illusions.

Main themes

• optical illusions
• diffusion models
• image generation
• AI art
• hardware requirements
• commercial use
• intellectual property
• entitlement and cost

What commenters say

• Some commenters think the technique is unrelated to a previously popular AI-generated image trend, while others see connections between the two.
• The difficulty of running the code and the need for high-end hardware are seen as barriers to experimenting with the technique by some, but others consider the cost to be relatively low.
• There is disagreement over whether the results of the technique constitute 'real art'.
• Some commenters are interested in exploring the potential of the technique for creating new types of images, such as recursive or hybrid images.
• The idea of a 'demixer' that can reverse-engineer the source images from a generated illusion is proposed as a potential area for further research.
• The entitlement of some individuals who expect to be able to use high-end hardware or software for free is criticized by others.
• The potential for commercial use of the technique is limited by the license of the underlying image generator, but it may be possible to obtain permission or find alternative generators.