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.
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.
This repository is currently a placeholder for the public release. We are organizing the codebase, benchmark registry, task specifications, and evaluation scripts for a clean release.
- Release the PhotoFlow agent code.
- Release the VPhotoBench scene registry and task specifications.
- Release benchmark construction notes and metadata.
- Release evaluation scripts for external metrics and aggregation.
- Release experiment configuration files for the reported baselines and ablations.
- Release selected logs and processed result tables.
- Add setup instructions for Blender, Python dependencies, and model/API configuration.
- Add example commands for running a single task and reproducing benchmark summaries.
VPhotoBench is built from publicly available Blender scenes. Some original assets may need to be downloaded from their upstream sources depending on license and file-size constraints. We will provide the scene registry, task metadata, preprocessing instructions, and scripts needed to reconstruct the benchmark layout.
The paper citation will be added after the public version is available.
The project license will be finalized before the full code release. Third-party Blender assets remain governed by their original licenses.