AI image generators may boast vast amounts of training data and promise limitless creativity, but it turns out that imagination is limited. New research shows that when forced to continually reinterpret their visuals, these models quickly abandon their originality and settle for some predictable, almost clichéd styles. The study, published in the journal Patterns, found that no matter how elaborate the opening prompt, AI image generators like Stable Diffusion XL and LLaVA end up surrounding the same dozens of artistic motifs. Think lighthouses, city skylines at night, rustic architecture, general indoor scenes, and other visuals that would be displayed in a hotel lobby rather than an art gallery.
Researchers have described this effect as “visual elevator music,” images that feel sophisticated but empty, not at all unpleasant, but ultimately forgettable. To test the AI's creativity, researchers designed something called the Visual Telephone Game. Setup was easy, at least in theory. One of the models, Stable Diffusion XL, was given a poetic prompt such as, “As I was sitting, especially alone, surrounded by nature, I found an old book, eight pages long, with a story told in a forgotten language, just waiting to be read and understood.”
Stable Diffusion generated an image from that text and passed it to another model, LLaVA, to explain it. The descriptions generated by LLaVA were sent back to Stable Diffusion, which used them to create new images. This process was repeated 100 times in a digital echo chamber, from image to description to new image.
Like a childhood game of telephone, messages became distorted as they passed from person to person, and the original concept quickly disappeared. By the 10th or 20th round, the resemblance to the initial image began to fade.
But it wasn't just the distortions that surprised the researchers, it was the convergence. After repeating the experiment 1,000 times, the model consistently gravitated toward 12 iteration styles. No matter how poetic, bizarre, or abstract the initial prompt, the end result always resembled one of the same motifs.
The researchers observed that stylistic changes usually occur gradually, with images losing their uniqueness one piece at a time. Sometimes the changes came suddenly and suddenly collapsed into blandness. But the destination was always the same. It was a general, glossy, faintly familiar sight.
The trends remained the same even when the team switched models using different versions of both the image generation and description tools. Extending the experiment to 1,000 turns only strengthened the pattern. By around the 100th sheet, the sequence stabilizes to a particular motif, with subsequent iterations producing only small variations on that theme.
“After repeating the telephone game 1,000 times, the researchers found that most image sequences ended up falling into one of 12 major motifs,” the study states. In other words, the machine had a comfort zone and liked to stay there.
So what does this say about artificial creativity? The researchers argue that while humans often introduce unpredictable interpretations, AI tends to smooth out irregularities. “In a human telephone game, each message is conveyed and heard differently, resulting in extreme variability,” the newspaper said. “AI has the opposite problem: it always defaults to a limited selection of styles, no matter how outlandish the original prompt.”
Of course, AI learns from human-generated data, but that can also contribute to this creative loop. When most of the photos online fall into similar categories like sunsets, streets, and interiors, it's no wonder that AI feels stuck playing familiar visual music.
– end
