Issues regarding prominent unnatural shapes have been pointed out in AI-generated food images used by restaurants and brands for advertising. Generated images include items such as stringy noodles, ice cream that resembles construction materials, and dishes riddled with countless holes.
Experts Explain Technical Reasons Why Generative AI Produces Unnatural Food Images
This article is a translation. Read the Japanese original
Experts point out that the cause lies in the mechanism of diffusion models, which most generative AI systems employ. According to Professor Chris Russell of the University of Oxford, diffusion models follow a process of gradually restoring detailed textures from noise.
During this process, the model may initially fail to correctly generate the basic structure of an object. The professor explains that in such cases, vivid textures are overwritten onto the incorrect structure, resulting in failures similar to images with an incorrect number of fingers.
Furthermore, there are weaknesses in rendering thin, continuous structures. Giovanni Battista Carifano of the University of Naples Federico II stated that noodles and string-like objects are geometric shapes that diffusion models struggle with. Consequently, a phenomenon occurs where textures encroach beyond logical boundaries.
These characteristics appear to be the factors producing unnatural patterns or hole-riddled images that can trigger trypophobia.
--- Source: Why AI food looks like that (The Verge AI, 2026-09-04)