DM3SR

DM3SR contains the scene-generation, rendering, question-answer generation, model-evaluation, and analysis code used for multi-view 3D spatial reasoning experiments.

Repository layout

• type1/, type2/, type3/: existing task-generation and evaluation code.
• type1b/: the finalized Type-1B pipeline and analysis release.
• eval/: shared model inference and scoring entry points.
• scripts/: experiment launch helpers.
• data/: compact benchmark and training annotations already maintained in Git.

The large Type-1B payloads are intentionally kept outside GitHub. Scene JSON, physics outputs, QA datasets, view-count ablations, model predictions, and rendered images are organized for a separate Hugging Face dataset release.

Type-1B release

See type1b/README.md for the canonical directory layout, dataset counts, pipeline commands, evaluation commands, and analysis workflow.

• GitHub payload: type1b/manifests/github_include.txt
• Hugging Face payload: type1b/manifests/huggingface_dataset_manifest.json
• Dataset card draft: type1b/manifests/HUGGINGFACE_DATASET_CARD.md
• Released dataset: Ever2after/DM3SR

Installation

python -m pip install -r requirements.txt

Model-specific dependencies may additionally be required by the external model repositories under eval/ and ViLaSR/.