CaptureGuide-Bench

CaptureGuide-Bench is the evaluation benchmark for ShutterMuse: Capture-Time Photography Guidance with MLLMs. It evaluates capture-time photography guidance from two complementary perspectives:

• Photographer-side / Composition Recommendation: given the current framing, decide whether it should be kept, refined, or rejected, and recommend composition boxes when refinement is needed.
• Subject-side / Pose Recommendation: given a scene image, recommend a scene-conditioned portrait pose. ShutterMuse predicts COCO-17 human keypoints with visibility states and can render the recommended pose for evaluation.

Links

Resource	Link
Paper	arXiv:2606.25763
Paper Page	Hugging Face Papers
Project Page	ShutterMuse Project
GitHub Repository	lijayuTnT/ShutterMuse
ShutterMuse Model	ShutterMuse/ShutterMuse

Dataset Structure

CaptureGuide-Bench/
├── photographer-side/
│   └── composition_benchmark/
│       ├── meta_new.json
│       └── original_composition/
└── subject-side/
    ├── paper-benchmark/
    └── paper-benchmark-gt/

Photographer-side: Composition Recommendation

• Task: evaluate composition decision and refinement quality.
• Images: photographer-side/composition_benchmark/original_composition/ contains 421 benchmark images.
• Annotations: photographer-side/composition_benchmark/meta_new.json stores image-level composition metadata, textual rationales, scores, and optional composition boxes.
• Main fields:
  • origin: rationale fields for the original framing, including good/bad reasons and composition type labels.
  • composition_boxes: candidate refinement boxes. Each rect is normalized as [x1, y1, x2, y2].
  • score: composition quality score used by the benchmark scripts.

Subject-side: Pose Recommendation

• Task: evaluate scene-conditioned portrait pose recommendation.
• Input images: subject-side/paper-benchmark/ contains 552 scene images.
• Reference images: subject-side/paper-benchmark-gt/ contains the paired reference/ground-truth images with matching file names.
• Evaluation: the ShutterMuse evaluation code compares generated pose visualizations against the paired references using VLM-based and pose-quality metrics.

Download

Download the full benchmark

hf download ShutterMuse/CaptureGuide-Bench \
  --repo-type dataset \
  --local-dir CaptureGuide-Bench

Download only the photographer-side split

hf download ShutterMuse/CaptureGuide-Bench \
  --repo-type dataset \
  --include "photographer-side/*" \
  --local-dir CaptureGuide-Bench

Download only the subject-side split

hf download ShutterMuse/CaptureGuide-Bench \
  --repo-type dataset \
  --include "subject-side/*" \
  --local-dir CaptureGuide-Bench

Download with Python

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="ShutterMuse/CaptureGuide-Bench",
    repo_type="dataset",
    local_dir="CaptureGuide-Bench",
)

If you use the official ShutterMuse evaluation scripts, place or symlink the downloaded benchmark as Benchmark/ under the ShutterMuse repository root:

git clone https://github.com/lijayuTnT/ShutterMuse.git
cd ShutterMuse
hf download ShutterMuse/CaptureGuide-Bench --repo-type dataset --local-dir Benchmark

The expected paths are:

ShutterMuse/Benchmark/photographer-side/composition_benchmark/meta_new.json
ShutterMuse/Benchmark/photographer-side/composition_benchmark/original_composition/
ShutterMuse/Benchmark/subject-side/paper-benchmark/
ShutterMuse/Benchmark/subject-side/paper-benchmark-gt/

Usage with ShutterMuse Evaluation

Install the project dependencies following the GitHub repository, then run one benchmark target at a time:

bash evaluation/scripts/run_unified_evaluation.sh photographer-model
bash evaluation/scripts/run_unified_evaluation.sh photographer-baseline
bash evaluation/scripts/run_unified_evaluation.sh subject
bash evaluation/scripts/run_unified_evaluation.sh subject-baseline

Common environment variables include OUTPUT_ROOT, PYTHON_BIN, model checkpoint paths, LoRA checkpoint paths, and API keys such as GEMINI_API_KEY, QWEN_API_KEY, and GPT_API_KEY for VLM-based scoring or API baselines.

Citation

If you use CaptureGuide-Bench, please cite:

@misc{li2026shuttermuse,
  title        = {ShutterMuse: Capture-Time Photography Guidance with MLLMs},
  author       = {Li, Jiayu and Fang, Yixiao and Hu, Tianyu and Cheng, Wei and Huang, Ping and Fan, Zheheng and Yu, Gang and Ma, Xingjun},
  year         = {2026},
  eprint       = {2606.25763},
  archivePrefix = {arXiv},
  primaryClass = {cs.CV},
  url          = {https://arxiv.org/abs/2606.25763}
}

License

The benchmark is released for research use. Please check the project repository and dataset repository for the latest license and usage terms.