AI Image Generators Fall Into Predictable Patterns

AI image generation models access vast visual datasets to create unique outputs. Yet, research reveals a surprising tendency: when repeatedly prompted wiht slightly altered requests, these models consistently revert to a limited set of visual themes, resulting in surprisingly generic imagery.

A study in Patterns tested Stable Diffusion XL and LLaVA using a visual telephone game. Stable Diffusion XL received an initial prompt-for exmaple, “As I sat particularly alone, surrounded by nature, I found an old book with exactly eight pages that told a story in a forgotten language waiting to be read and understood.” It generated an image,which LLaVA then described.That description became the new prompt for Stable Diffusion, and the process repeated for 100 rounds.

As with human telephone, the original image quickly degraded. This isn’t shocking. You’ve likely seen time-lapse videos demonstrating how AI models struggle to reproduce an image without alteration,only for