The use of AI-generated images by restaurants, cafés and brands in food promotions is producing unappetizing results such as doughnut-shaped shrimp, worm-like noodles, stringy chicken, rock-like burgers and burritos covered in holes. The distortions in these images reveal the technical limitations of AI systems in image generation and problems in their training data.
Most leading image generators use diffusion models. These systems create images by starting with pure noise and gradually reducing the noise; first, a rough structure emerges, followed by fine details. According to Chris Russell from University of Oxford, when the model incorrectly establishes an object’s basic form at an early stage, it can later add vivid textural details onto the flawed structure. This is likened to creating a person with six fingers.
Giovanbattista Califano from University of Naples Federico II says diffusion models are weak at generating thin, continuous structures that end at a point. As a result, noodles, fibers and curved extensions can be shifted to the wrong places; repeated textures such as bubbles and seeds can also extend beyond logical boundaries. The excessively noodle-like forms and clusters of holes in the images can trigger the fear of holes known as trypophobia.
According to experts, AI does not truly know what a sandwich, noodle or burrito is. It only statistically learns what these objects generally look like on the internet; it does not understand their functions and behavior in the physical world. Roland Meyer from University of Zurich says the systems imitate the appearance of a photograph but do not understand the context behind professional photographic aesthetics. As a result, ice cream can look like concrete, while a burger can look like stone.
Training data can also exacerbate the problem. The bright lighting, intense colors and exaggerated shapes in food photographs are sometimes transferred to the model along with foods that do not exist in reality. Simon Colton from Queen Mary University of London notes that strange and viral content attracts more attention on the internet than ordinary images. In addition, AI-generated content training new models can lead to visual distortion and convergence through “model collapse.”
Ambiguous prompts, the enlargement of low-resolution images and inadequate system instructions make errors more visible. People, meanwhile, quickly notice abnormalities in food because of the sense of disgust that evolved in response to the dangers of parasites, contamination and spoilage. As a result, AI-generated food looks not only unrealistic but also instinctively wrong.
Why it matters
This problem shows that AI-prepared food promotions are not merely an aesthetic choice; distortions in the image can directly affect whether viewers find the product appetizing or trustworthy. For restaurants, cafes and brands, the issue is the need to subject the results to human review because they emerge from the model combining similar images without grasping the object’s physical structure. Details such as noodles, fibers, bubbles and holes in particular make technical errors visible and can also trigger a disgust response in some viewers. The reintroduction of strange and AI-generated images from the internet into training data leaves open the question of whether distortions will be repeated.