README.md

metadata

pretty_name: VPhotoBench
license: other
language:
  - en
size_categories:
  - 100<n<1K
tags:
  - blender
  - 3d
  - virtual-photography
  - vision-language-models
  - camera-control
  - agentic-ai
configs:
  - config_name: tasks
    data_files:
      - split: train
        path: benchmark/tasks.csv
  - config_name: scenes
    data_files:
      - split: train
        path: benchmark/scenes.csv
  - config_name: attribution
    data_files:
      - split: train
        path: licenses/ATTRIBUTION.csv
  - config_name: licenses
    data_files:
      - split: train
        path: licenses/LICENSES_PER_ASSET.csv
  - config_name: benchmark_list
    data_files:
      - split: train
        path: metadata/benchmark_list_clean.csv

VPhotoBench

VPhotoBench is a language-conditioned virtual photography benchmark built from 47 open-license Blender scenes. Each scene is paired with three photography missions: subject placement, relational composition, and atmosphere/style. This release contains 47 Blender .blend scene files and 141 English task prompts.

Contents

• Scenes: 47
• Tasks: 141
• Mission types: subject placement, relational composition, atmosphere/style
• Total Blender asset size: 6.34 GiB
• Per-asset licenses: CC-BY: 22, CC-BY-ND: 1, CC-BY-SA: 12, CC0: 12

Directory structure

assets/blender/                 # 47 Blender scene files, one directory per scene
benchmark/scene_registry.json   # Scene metadata and package-relative file paths
benchmark/scene_registry_v1.json # Runner-compatible scene registry
benchmark/tasks.json            # 141 task prompts
benchmark/tasks_v1.json         # Runner-compatible task registry
benchmark/split_v1.json         # Canonical dev/test split used by the paper scripts
benchmark/tasks.csv             # Flat task table
evaluation/EVALUATION_PROTOCOL.md # How to score rendered images with UniPercept
evaluation/run_unipercept_scoring.py # UniPercept-based image scoring helper
metadata/resource_manifest.json # File sizes, checksums, and package paths
metadata/checksums_sha256.txt   # SHA256 checksums for all packaged .blend files
metadata/benchmark_list_original.csv  # Original source table
metadata/benchmark_list_original.xlsx # Original source spreadsheet, when present
licenses/ATTRIBUTION.csv        # Per-scene TASL-style attribution table
licenses/ATTRIBUTION.md         # Human-readable attribution summary
licenses/LICENSES_PER_ASSET.csv # Per-scene license table
scripts/validate_package.py     # Local package integrity checker

Usage

The task files use package-relative paths. A task entry points to a scene_id; the corresponding scene entry gives the .blend file path under assets/blender/.

import json
from pathlib import Path

root = Path("VPhotoBench")
scenes = {s["scene_id"]: s for s in json.loads((root / "benchmark/scene_registry.json").read_text())["scenes"]}
tasks = json.loads((root / "benchmark/tasks.json").read_text())["tasks"]

task = tasks[0]
blend_file = root / scenes[task["scene_id"]]["blend_file"]
print(task["task_id"])
print(blend_file)
print(task["instruction"])

JSON files are the authoritative task and scene registries. CSV files are provided for convenience and are encoded as UTF-8.

Evaluation

VPhotoBench does not require a built-in evaluator inside the Blender assets. After a method renders final images for the task prompts, users can score those images with UniPercept:

https://github.com/thunderbolt215/UniPercept

We provide a small helper in evaluation/run_unipercept_scoring.py. Prepare a CSV file with task_id,image_path, then run:

See evaluation/EVALUATION_PROTOCOL.md for details. The helper reports UniPercept IAA, IQA, and ISTA scores, plus an optional default composite M_qs = 0.40 * IAA + 0.20 * IQA + 0.40 * ISTA.

Licensing and attribution

The benchmark is a metadata and task release over third-party Blender assets. The dataset as a whole does not impose a single blanket license over all assets. Each .blend file remains under its upstream license. See:

• licenses/LICENSES_PER_ASSET.csv
• licenses/ATTRIBUTION.csv
• licenses/ATTRIBUTION.md

For Creative Commons assets, this release follows a TASL-style attribution table where available: title, author, source URL, and license. Some upstream entries do not have a stable source URL in the local project records; these rows are explicitly marked in attribution_status. Current unresolved or partial attribution rows: 0.

Users are responsible for checking each upstream license before redistribution or commercial use. In particular, rows marked as fan art or partial attribution should be reviewed against the upstream source.

Dataset construction

The 47 scenes were selected from public Blender scene repositories and paired with three natural-language photography instructions per scene. Each instruction asks an agent to choose an executable camera state in the original 3D scene rather than generate a new image.

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},
}