PhotoFlow: Agentic 3D Virtual Photography Missions

PhotoFlow is a closed-loop agent for virtual photography in Blender scenes, using text prompts to propose, judge, and revise camera shots.

It combines a Director-Reviewer-Reflector workflow to balance 3D scene understanding with photographic aesthetics.

The authors also introduce VPhotoBench, a benchmark with 47 scenes and 141 photography missions.

On held-out tests, PhotoFlow outperformed several baselines under a limited rendering budget, suggesting LLM-based agents can support embodied visual planning and creative image generation.