GenEvolve: Self-Evolving Image Generation Agents via Tool-Orchestrated Visual Experience Distillation

Researchers introduced GenEvolve, a self-improving image generation framework that treats each attempt as a tool-driven trajectory.

It compares multiple generation attempts for the same request and distills structured visual experience into teacher-student training.

The system improves prompt construction, reference selection, and composition for more effective image generation.

GenEvolve achieved stronger results on public benchmarks and a new GenEvolve-Bench, showing gains over one-shot prompting.