STRANGE-Bench

STRANGE-Bench measures typicality bias: how much a vision model's recognition of an object depends on the objects it usually appears with. It is the benchmark of the NeurIPS 2026 paper Stranger Things: When Objects Appear Without Their Typical Neighbors (paper (link to be added); code: https://github.com/sidgairo18/stranger-things).

Each scene is a photorealistic 1024 x 1024 render of an anchor object from the 80 COCO categories together with k partner objects. The partners are chosen by their normalized pointwise mutual information (NPMI) with the anchor in COCO train2017: typical partners (high NPMI), atypical partners (low NPMI), or partners that never co-occur with the anchor there. At each object count, the typical and atypical sub-tiers hold the same number of scenes, so the gap between them measures the effect of context with the number of objects held fixed.

Intended use: evaluation only. Please do not train on STRANGE-Bench: training on its atypical scenes removes the contrast it measures. This is a request; the license (CC BY 4.0) does not restrict use.

At a glance

	scenes	instance annotations	ignore regions
balanced/: benchmark of record (every number in the paper)	5,523	17,332	300
clean/: before balancing, included for transparency	6,571	21,349	368
samples/: inspection sample of balanced/	170

heldout_class_scene_ids.txt lists the 3,234 scenes of balanced/ in which no class appears in the generated training scenes of the paper's AAS late stage (the held-out-class subset of App. A.6 of the paper).

Sub-tiers

k is the number of partners next to the anchor. The paper's main tables use these eleven sub-tiers:

sub-tier	k	partner type	balanced scenes	balanced instances	clean scenes	held-out-class scenes
tier1_natural_bg	0	none, natural background	421	421	422	421
tier1_atypical_bg	0	none, unusual background	421	421	448	421
tier2_1typical	1	typical (high NPMI)	400	800	403	327
tier2_1atypical	1	atypical (low NPMI)	400	800	412	329
tier3_2typical	2	typical	392	1,176	435	199
tier3_2atypical	2	atypical	392	1,176	414	219
tier4_3typical	3	typical	399	1,596	459	124
tier4_3atypical	3	atypical	399	1,596	465	158
tier5_4typical	4	typical	229	1,145	360	50
tier5_4atypical	4	atypical	229	1,145	313	64
tier_never	1-2	never co-occurring	466	1,329	507	314

Six further sub-tiers are reported in the paper's appendix:

sub-tier	k	partner type	balanced scenes	balanced instances	clean scenes	held-out-class scenes
tier2_1typical__small_anchor	1	typical, small anchor	200	400	204	170
tier2_1atypical__small_anchor	1	atypical, small anchor	200	400	261	163
tier6_5typical	5	typical	193	1,158	432	17
tier6_5atypical	5	atypical	193	1,158	332	26
tier_never_k3	3	never co-occurring (clique)	334	1,336	381	165
tier_never_k4	4	never co-occurring (clique)	255	1,275	323	67

The tier1_* scenes show the anchor alone, on a natural or on an unusual background. tier_never pairs the anchor with one or two partners that never co-occur with it in COCO train2017; tier_never_k3 and tier_never_k4 are cliques of three and four partners in which no pair co-occurs. The __small_anchor sub-tiers repeat k = 1 with a small anchor.

Files

balanced/<sub_tier>/images/<scene_id>.png        1024 x 1024 PNG
balanced/<sub_tier>/annotations.json             COCO instance annotations (boxes and SAM3 RLE masks)
balanced/<sub_tier>/annotations_complete.json    scenes in which every intended class is annotated
clean/<sub_tier>/...                             same layout, before balancing
samples/<sub_tier>/{images/, annotations.json}   10 scenes per sub-tier of balanced/ (seed 0)
heldout_class_scene_ids.txt                      held-out-class subset, one scene id per line
scene_metadata.jsonl                             one record per scene of clean/: spec, planner layout, caption, seed
tier_manifest.csv                                one row per scene: sub-tier, anchor, partners, k, membership flags
stats.csv                                        counts per sub-tier
categories.json                                  COCO-80 index, name and COCO id
croissant.json                                   Croissant metadata with Responsible-AI fields
code/                                            generation code of this release (code/README.md)
LICENSE-DATA, LICENSE-CODE, SHA256SUMS

Every scene of balanced/ is complete; the paper evaluates annotations.json. Scene ids start with v2_, the name of the internal generation batch; the prefix carries no version meaning.

Annotation format

COCO instance format, one file per sub-tier, with the 80 COCO categories and their original ids.

• images: id, file_name, width, height, scene_id, anchor, labels_intended (the classes requested for the scene).
• annotations with iscrowd = 0: one instance per intended class. bbox (x, y, w, h in pixels) is the bounding box of the SAM3 mask in segmentation (COCO RLE); score is the SAM3 confidence.
• annotations with iscrowd = 1: ignore regions, unintended objects found by the second audit (SAM3 and OWLv2 agree). Standard COCO evaluation (pycocotools) does not count detections on them as false positives. They are kept in typical scenes, or where the object's NPMI with every intended class is at most 0; an atypical scene with any other unintended object was dropped, so no unlabeled typical partner sits in an atypical scene.

How the scenes were made

1. Layout and caption (Qwen2.5-3B-Instruct). For each scene, k + 1 categories are sampled for the target sub-tier. The language model writes a bounding-box layout (a rule-based solver from COCO size priors is the fallback, and the only planner for k = 5 and for the never-co-occurring cliques) and a one-sentence photographic caption that names each object with its position and a short visual description. 342 scenes of balanced/ use the deterministic template caption that replaces a failed language-model call.
2. Rendering (Stable Diffusion 3.5 Large). Up to three renders per scene (40 steps, guidance 3.5) under a fixed negative prompt against crops, clutter and duplicates.
3. Segmentation (SAM3). One text prompt per intended class ("a {label}", with overrides for ambiguous COCO names); the first render in which SAM3 finds every intended class is kept.
4. Consistency audit (Grounding DINO Swin-B). A scene with an extra class (score at least 0.5) or a duplicate of an intended class is dropped.
5. Second audit and balancing (SAM3 and OWLv2). Unintended objects found by both models (SAM3 score at least 0.5, OWLv2 score at least 0.2, box IoU at least 0.3) become ignore regions or drop the scene (see above); scenes that miss an intended class are dropped; the typical and atypical sub-tiers at each k are cut to the same size by anchor-matched subsampling. Of the rendered scenes, 44.6 % enter the benchmark.

scene_metadata.jsonl records, per scene, the planner layout, the caption, the render seed and the number of renders.

code/ holds the code of every stage with the settings of this release; code/README.md gives the recipe and what was checked. On the stored intermediates of the build it reproduces the images, boxes, ignore regions and masks of clean/ and balanced/ exactly, and released scenes re-render byte-identically from their stored captions and seeds (spot check on the GPU type of the build).

Properties to keep in mind

These follow from the construction and are discussed in App. C.5 of the paper.

• The atypical sub-tiers are not limited to low-NPMI pairs: pairs that never co-occur in COCO train2017 count as maximally atypical, so 39 to 64 % of their anchor-partner pairs never co-occur there. In 76 of the 466 tier_never scenes the anchor co-occurs with almost every class (for example person, car, cup), and the sampler takes the anchor's rarest partners instead; 40 of these scenes contain no never-co-occurring pair.
• When no candidate passes the typical rule, the sampler takes the best-scoring one, so 3.5, 10.3, 16.5 and 20.6 % of the typical anchor-partner pairs at k = 1 to 4 have NPMI below 0.1.
• The caption request for the atypical sub-tiers with partners used the wording of the typical ones, so the two arms differ in the requested classes, not in the context wording. The final captions mention an unusual context in 6 to 16 % of the atypical scenes and in none of the typical ones.
• Boxes are the bounding boxes of the SAM3 masks; no audit compares them with the planner layout.

Evaluating a model

Report COCO box AP per sub-tier on annotations.json, the gap AP(typical) - AP(atypical) at each k, and the drop AP(tier2_1typical) - AP(tier_never). Classifiers can be scored on the label set of each image. The evaluation scripts of the paper are in the code repository (https://github.com/sidgairo18/stranger-things).

Human validation

Three blinded studies in the paper (App. C.7) used an initial version of the benchmark built with the same layout, rendering and segmentation stages but filtered by the Grounding DINO audit alone. Human raters judged 93.4 % of its scenes photorealistic, identified the anchor in 99.5 % and the boxes as correct in 98.2 % (COCO photographs with their own annotations: 99.7, 98.5 and 98.3 %). Human anchor recall dropped by 1.0 point from tier1_natural_bg to tier_never scenes, where a fine-tuned DINOv3 classifier lost 31.4 points.

Limitations and biases

• The scenes are synthetic: composition, lighting, textures and the appearance of people follow Stable Diffusion 3.5 Large. They are minimalist wide shots and easier to read than COCO photographs with the same number of objects.
• Typicality is defined by COCO train2017 co-occurrence statistics. Classes that are rare in COCO are rare here, and NPMI measures statistical co-occurrence, not semantic plausibility.
• The vocabulary is COCO-80 only, and mask quality varies by class.
• No image shows a real person, and the dataset holds no personal data.

License

CC BY 4.0 (see LICENSE-DATA). The generation code in code/ is under the MIT license (see LICENSE-CODE). The images were generated with Stable Diffusion 3.5 Large (Stability AI Community License); the categories and the co-occurrence statistics come from the COCO 2017 annotations (CC BY 4.0).

Citation

@inproceedings{gairola2026stranger,
  title     = {Stranger Things: When Objects Appear Without Their Typical Neighbors},
  author    = {Gairola, Siddhartha and Xie, Jiahao and Kukleva, Anna and Locatello, Francesco and Schiele, Bernt},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Evaluations and Datasets Track},
  year      = {2026}
}

sha256sum -c SHA256SUMS verifies a download.