AI Image Generators Default to the Same 12 Photo Styles, Study Finds - Gizmodo

AI image generators may have vast training data, but a new study finds they keep circling back to the same looks. Researchers report that when models iterate on their own outputs in a “visual telephone” loop, they converge on just a dozen generic photo styles.

What the researchers did

• Models tested: Stable Diffusion XL (image generation) and LLaVA (image description)
• Method: SDXL created an image from a short prompt. LLaVA described that image. The description was fed back to SDXL to create the next image. This cycle repeated for 100 rounds.
• Scale: 1,000 separate runs of the telephone game; additional tests swapped in other image and captioning models, and some runs extended to 1,000 turns.

Key findings

• Rapid drift: Like the party game telephone, the original concept quickly vanished as images morphed from round to round.
• Style collapse: Across most runs, sequences settled into one of 12 dominant visual motifs—the researchers dubbed this “visual elevator music,” akin to bland hotel art.
• Common motifs: Maritime lighthouses, formal interiors, urban night scenes, rustic architecture, and similar broadly appealing but generic styles.
• Timing and persistence: Convergence typically happened by around turn 100. Extending runs to 1,000 turns produced variations, but they still stayed within those popular motifs.
• Robust across models: Even when the team swapped in different image and captioning models, the same convergence patterns emerged.

Why it matters

• Creativity limits: Despite massive training sets, the systems default to a narrow visual vocabulary, suggesting constrained generative diversity under iterative refinement.
• Data bias echoes: These defaults likely mirror what’s overrepresented and aesthetically reinforced in training data and human-made captions.
• Big picture: Copying style is easy; curating taste—and sustaining originality over iterative generations—is much harder for current AI.

Bottom line
AI image generators can produce striking pictures, but when asked to reinvent themselves repeatedly, they tend to settle into safe, familiar looks. The study in the journal Patterns highlights how today’s models, left to their own feedback loops, gravitate toward 12 predictable styles rather than open-ended creativity.

Source: https://gizmodo.com/ai-image-generators-default-to-the-same-12-photo-styles-study-finds-2000702012