code/README.md

STRANGE-Bench generation code

The code that generated this dataset from the COCO train2017 annotations and public models: co-occurrence statistics, scene specifications, rendering, automatic audits, the clean/ and balanced/ trees, and instance masks. The launcher defaults are the settings of this release.

Layout

code/
├── README.md, requirements.txt, env_template.sh
├── launchers/          one script per stage (common.sh holds the shared paths)
├── strangebench/
│   ├── cooccurrence/   NPMI statistics and partner sampling
│   ├── scene_specs/    sub-tiers, anchor and partner classes, layouts
│   ├── generation/     captioning, rendering and segmentation; specs_from_metadata.py
│   ├── audit/          detector audit and the clean tree
│   ├── strict/         second audit, strict filter, clean / balanced build
│   ├── masks/          instance masks
│   └── common/         shared helpers
└── tests/              CPU tests (pytest)

Setup

Two Python environments, listed with their versions in requirements.txt: a generation environment (PyTorch, diffusers, transformers, SAM 3) and an MMDetection 3.x environment for the Grounding DINO audit. Paths, the Hugging Face cache and the detector checkpoint are set in env_template.sh.

Models: Qwen/Qwen2.5-3B-Instruct (layouts, captions), stabilityai/stable-diffusion-3.5-large (rendering; gated, accept its license on the Hub first), SAM 3 (facebook/sam3; segmentation, audit, masks), Grounding DINO Swin-B from MMDetection (audit) and google/owlv2-large-patch14-ensemble (audit). No weights are bundled.

Running

Edit and source env_template.sh, activate the environment of each stage, and run the launchers in order. Sharded stages take SHARD_ID (or SLURM_ARRAY_TASK_ID); each launcher has an optional Slurm header. Outputs go under $WORK_ROOT.

stage	launcher	GPU	what it does
0	stage0_cooccurrence.sh	no	co-occurrence statistics (NPMI) from COCO train
1	stage1_scene_specs.sh	yes	scene specifications: sub-tier, anchor, partners, layout
2	stage2_render.sh	yes	caption, render (up to 3 attempts) and segment each scene
3	stage3_lean_audit.sh	yes	Grounding DINO audit for unrequested and duplicate objects
4	stage4_clean_tree.sh	no	apply the audit and build the clean tree
5	stage5_independent_audits.sh	yes	SAM 3 and OWLv2 detections of all 80 classes
6	stage6_strict_build.sh	no	strict filter, then the clean/ and balanced/ trees
7	stage7_masks.sh	yes	instance masks (COCO RLE) for the released trees

The full run takes about 200 GPU-hours, most of it in stage 2. Tests: cd code && pytest tests.

Re-rendering released scenes

python code/strangebench/generation/specs_from_metadata.py --metadata scene_metadata.jsonl --out specs.json \
    [--scene-ids ids.txt]
RAW_ROOT=$WORK_ROOT/rerender SCENE_SPECS=specs.json CAPTION_FROM_SPEC=1 NUM_SHARDS=1 bash code/launchers/stage2_render.sh

Each scene is rendered from its stored caption with the seeds of the release.

Notes

• Layouts and captions are sampled from the language model, so a new run gives new ones; the released ones are in scene_metadata.jsonl.
• Use numpy 2 for stage 1 to reproduce the released scene specifications.
• LEGACY_CAPTION_WORDING=1 in stage 2 uses the caption requests of this release, in which the atypical sub-tiers were requested with the typical context wording; the default uses the wording described in the paper.

License

MIT, see LICENSE-CODE at the root of this repository. The dataset is CC BY 4.0 (LICENSE-DATA).