Autonomous language-image generation loops converge to generic visual motifs

Area of Science:

• Artificial Intelligence
• Machine Creativity
• Human-Computer Interaction

Background:

• AI-to-AI creative systems offer novel avenues for machine creativity.
• Concerns exist regarding the potential for these systems to generate homogenous outputs.

Purpose of the Study:

• To investigate the convergence tendencies of autonomous AI-to-AI creative systems.
• To analyze the diversity and motifs generated by iterated cross-modal AI feedback loops.

Main Methods:

• Developed iterative feedback loops between Stable Diffusion XL (image generation) and LLaVA (image description).
• Executed 700 trajectories across diverse prompts and 7 temperature settings over 100 iterations.
• Quantitatively analyzed visual outputs for dominant motifs and aesthetic patterns.

Main Results:

• All AI-to-AI creative system runs converged to highly similar, generic visuals, termed 'visual elevator music'.
• Identified 12 dominant motifs characterized by commercially safe aesthetics (e.g., stormy lighthouses, palatial interiors).
• Convergence persisted across different model pairs, suggesting structural limitations in cross-modal AI.

Conclusions:

• Autonomous AI-to-AI systems exhibit inherent homogenizing tendencies, collapsing creative diversity.
• This convergence mirrors human cultural transmission biases but results in complete diversity loss.
• There is a critical need for anti-convergence mechanisms and human-AI collaboration to foster genuine creative diversity.