PhotoFlow: Agentic 3D Virtual Photography Missions

PhotoFlow is an agentic framework for language-conditioned virtual photography in controllable 3D scenes. Given a Blender scene and a natural-language photography intent, PhotoFlow searches for an executable camera state, including camera pose, look-at target, lens, aperture, and aspect ratio, then renders the final photograph.

Teaser

Overview

PhotoFlow treats virtual photography as a closed-loop spatial-aesthetic search problem. The agent is organized around three roles:

• Director proposes diverse candidate camera states from scene scouts, soft photographic blueprints, global anchors, and region memory.
• Reviewer evaluates rendered previews using structured rule checks, visual critique, and pairwise incumbent selection.
• Reflector converts failures into search bias, dead-zone suppression, and high-exploration relocation.

We also introduce VPhotoBench, a benchmark of language-conditioned virtual photography missions over open-license Blender scenes. The benchmark is designed to evaluate whether an agent can satisfy spatial constraints, semantic intent, aspect-ratio choices, and photographic quality in fully virtual art scenes.

Resources

• Paper: arXiv:2605.23771
• Hugging Face paper page: huggingface.co/papers/2605.23771
• Project page: visionary-laboratory.github.io/PhotoFlow
• Benchmark dataset: VPhotoBench on Hugging Face

Release Status

• Agent Code Director-Reviewer-Reflector implementation, prompts, JSON schemas, and run configurations.
• VPhotoBench Scene registry, task specifications, benchmark construction notes, and asset metadata.
• Evaluation External metric aggregation, baseline configs, ablation summaries, and selected logs.

Benchmark Assets

VPhotoBench is built from publicly available Blender scenes and contains 47 Blender scene files paired with 141 language-conditioned photography missions. Third-party Blender assets remain governed by their original licenses; see the Hugging Face dataset card for per-asset license and attribution metadata.

Citation

@misc{guo2026photoflowagentic3dvirtual,
      title={PhotoFlow: Agentic 3D Virtual Photography Missions},
      author={Jiarui Guo and Haojia Wei and Yiming Zhang and Yifei Liu and Yuning Gong and Hongjie Zhang and Xue Yang and Zhihang Zhong},
      year={2026},
      eprint={2605.23771},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2605.23771},
}

License

The project license will be finalized before the full code release. Third-party Blender assets remain governed by their original licenses.