AI image generation models have massive sets of visual data to pull from in order to create unique outputs. And yet, researchers find that when models are pushed to produce images based on a series of slowly shifting prompts, it’ll default to just a handful of visual motifs, resulting in an ultimately generic style.
A study published in the journal Patterns took two AI image generators, Stable Diffusion XL and LLaVA, and put them to test by playing a game of visual telephone. The game went like this: the Stable Diffusion XL model would be given a short prompt and required to produce an image—for example, “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.” That image was presented to the LLaVA model, which was asked to describe it. That description was then fed back to Stable Diffusion, which was asked to create a new image based off that prompt. This went on for 100 rounds.
Much like a game of human telephone, the original image was quickly lost. No surprise there, especially if you’ve ever seen one of those time-lapse videos where people ask an AI model to reproduce an image without making any changes, only for the picture to quickly turn into something that doesn’t remotely resemble the original. What did surprise the researchers, though, was the fact that the models default to just a handful of generic-looking styles. Across 1,000 different iterations of the telephone game, the researchers found that most of the image sequences would eventually fall into just one of 12 dominant motifs.
In most cases, the shift is gradual. A few times, it happened suddenly. But it almost always happened. And researchers were not impressed. In the study, they referred to the common image styles as “visual elevator music,” basically the type of pictures that you’d see hanging up in a hotel room. The most common scenes included things like maritime lighthouses, formal interiors, urban night settings, and rustic architecture.
Even when the researchers switched to different models for image generation and descriptions, the same types of trends emerged. Researchers said that when the game is extended to 1,000 turns, coalescing around a style still happens around turn 100, but variations spin out in those extra turns. Interestingly, though, those variations still typically pull from one of the popular visual motifs.
So what does that all mean? Mostly that AI isn’t particularly creative. In a human game of telephone, you’ll end up with extreme variance because each message is delivered and heard differently, and each person has their own internal biases and preferences that may impact what message they receive. AI has the opposite problem. No matter how outlandish the original prompt, it’ll always default to a narrow selection of styles.
Of course, the AI model is pulling from human-created prompts, so there is something to be said about the data set and what humans are drawn to take pictures of. If there’s a lesson here, perhaps it is that copying styles is much easier than teaching taste.
Read the full article here
