AI Image Generators Default to the Same 12 Photo Styles, Study Finds (gizmodo.com)

A recent study published in the journal Patterns reveals that popular AI image generators, specifically Stable Diffusion XL and LLaVA, tend to converge on a limited set of visual styles—just 12 dominant motifs—when tasked with creating images over a series of evolving prompts. In a creative twist on the classic telephone game, researchers observed that as these models produced and reinterpreted images iteratively, they ultimately regressed to generic representations, often resembling common scenes like urban night settings and maritime lighthouses. This phenomenon occurred consistently across 1,000 iterations, highlighting a striking lack of creativity in AI systems. This research is significant for the AI/ML community as it underscores potential limitations in the creativity of generative models, suggesting they are significantly influenced by the biases inherent in their training datasets. The term “visual elevator music” aptly characterizes these generic outputs, pointing to an issue where unique prompts fail to elicit equally unique visual responses. As AI systems default to familiar styles rather than exhibit true innovation, the findings prompt a reevaluation of how creativity is defined and realized in AI, emphasizing the challenge of teaching AI models a taste for variety over mere replication.