Overview
- A Patterns paper by Hintze et al. tested iterative handoffs between Stable Diffusion XL and LLaVA across 100 diverse prompts to probe autonomous creativity.
- Image–description loops drifted off the original prompts and repeatedly converged on 12 recurring themes, including lighthouses, urban night scenes, rustic architecture, formal interiors, and sports imagery.
- The convergence emerged across different generator and describer pairings, longer prompts, and higher randomness settings, indicating robust behavior.
- Sequences typically stabilized by about 100 back‑and‑forth iterations, and in runs extended to 1,000 iterations they sometimes jumped to a different generic motif after long stable periods.
- The authors describe the outcomes as “visual elevator music,” attribute them to training‑data biases and weak evaluative judgment in current models, and recommend anti‑convergence techniques plus human‑in‑the‑loop workflows.