Recent research published in the journal Patterns reveals the creative limitations of AI image generators like Stable Diffusion XL and LLaVA. Despite extensive training data, these models often default to a limited repertoire of clichéd styles—12 recurring motifs, including lighthouses and city skylines. Researchers likened the generated visuals to “visual elevator music,” describing them as polished yet forgettable. The study involved a digital “telephone game,” where images were repeatedly reinterpreted, leading to increasingly generic results while diminishing the original concept. Regardless of the initial prompt’s complexity, the final images gravitated toward familiar themes. This pattern persisted even with different model iterations, suggesting that AI tends to smooth out creative irregularities unlike humans, who introduce diverse interpretations. The results underscore the implications of AI’s dependence on common visual data, limiting its creative potential and reinforcing its tendency to reproduce familiar visuals over originality.
Share
Read more