AI image generation is a fascinating field, but recent research shows it may not be as creative as we hope. A study in the journal Patterns examined how two AI models, Stable Diffusion XL and LLaVA, created images based on shifting prompts. They played a kind of visual telephone game: one model generated an image from a prompt, the other described it, and then the first model created a new image from that description. They repeated this process for 100 rounds.
The outcome? Much like the childhood game, the original image was lost along the way. Surprisingly, researchers found that the models often defaulted to just a few common styles, which they dubbed “visual elevator music.” Most of the images ended up resembling familiar scenes like lighthouses, urban nights, and rustic buildings.
Researchers ran 1,000 iterations of this experiment, and in almost all cases, the image style coalesced around 12 dominant motifs by the 100th turn. Even when they switched models, the same trends appeared. After 1,000 turns, while some variations appeared, they still remained close to those popular styles.
This behavior highlights a key limitation of AI: it has a hard time being genuinely imaginative. In a human game of telephone, each person filters the message through personal biases and preferences, creating a wide range of interpretations. AI, however, often falls back on a narrow selection of styles, reflecting the limitations of its training data.
This isn’t solely the AI’s fault; it’s also a reflection of what people typically capture and label in images. According to a 2022 Statista survey, over 70% of Instagram posts feature familiar themes like nature or urban life. This trend suggests that the models are primarily trained on data that prioritize common, appealing images.
The broader implication? Teaching AI true creativity may require more than just algorithms; it’s about understanding human taste and diversity in expression. So, while AI can produce stunning visuals, fostering true innovation may still be a ways off.
Source link
Artificial intelligence,emerging technologies,generative ai,generative artificial intelligence,image generation model
