CompoSET — Compositional Single-Edit Testbed

CompoSET is a vision-language compositionality benchmark built on the principle of same scene, one edit. Each pair of images (base, var) differs by exactly one localized attribute change — e.g. color, material, pattern, pose, spatial arrangement — while the rest of the scene is held constant. This lets evaluators attribute model errors to the specific compositional edit, isolated from broader scene-composition confounds that plague prior benchmarks.

Dataset summary

• 80 scenes, each containing multiple objects and — in many cases — people.
• 1,776 live variations (one per base / foil image pair), grouped by scene.
• 16 edit types in the edit_types_v2 taxonomy (12 modify_* + 4 swap_*).
• Three caption density tiers per variation: short (ARO-style), medium (COCO-style), long (Flickr30k-style).
• Single evaluation split: test (no train split; this is a benchmark).

Dataset structure

The release ships three parquet files and a flat image directory. See RELEASE_LAYOUT.md in the source repo for a working cheatsheet (GitHub URL TBD).

variations.parquet — 1,776 rows, canonical eval unit

column	type	description
id	str	"s001_v01"
scene_id	str	"s001"
edit_type	str	one of 16 categories (see below)
image_base	str	path relative to repo root, e.g. "images/s001_base.png"
image_var	str	path, e.g. "images/s001_v01.png"
captions	struct	{short, medium, long: {base, var}} — see styling notes

Caption styling (matches each tier's reference benchmark style):

• short: lowercase noun phrase, no terminal period (e.g. "a red metal tin").
• medium: full sentence, Capital + period (e.g. "A red metal tin on the right side of the cart.").
• long: full scene sentence, Capital + period.

Each variation defines an unambiguous pair: base_image ↔ base_caption, var_image ↔ var_caption. No scrambling, no separate answer key.

scenes.parquet — 80 rows, per-scene metadata

column	type	description
scene_id	str	"s001"
image_base	str	path to the scene's base image
setting	str	free-text scene description, e.g. "dentist's office examination room"
setting_category	str	"indoor" (41 scenes) or "outdoor" (39 scenes)
n_subjects	str	"0" (25) / "1" (23) / "2" (16) / "3+" (16)
framing	str	"close-up" (30) / "medium" (23) / "full" (27)
n_variations	int	number of live variations in this scene
edit_types_present	list[str]	subset of the 16-type taxonomy present in this scene, in canonical order

scene_captions.parquet — 80 rows, reference / provenance

column	type	description
scene_id	str	"s001"
base_caption	str	full corrected caption for the scene (one paragraph, ~4–6 sentences), used to drive image generation and the long density tier

images/ — 1,856 PNGs

• s{NNN}_base.png — one per scene (80 files).
• s{NNN}_v{XX}.png — one per live variation (1,776 files).
• Variations are renumbered contiguously within each scene (v01..vNN), so there are no gaps. The original (internal) IDs are available in the source repo for provenance.

Edit type taxonomy

16 categories grouped by semantic axis. Canonical ordering (used in tables and figures):

edit_type	n	axis
modify_color	206	visual attribute
modify_pattern	101	visual attribute
modify_shape	103	visual attribute
modify_material	161	visual attribute
modify_transparency	101	visual attribute
modify_spatial	99	spatial relation
modify_state	146	state / quantity
modify_presence	104	state / quantity
modify_cardinality	100	state / quantity
modify_action	101	action / pose
modify_pose	104	action / pose
modify_object	101	object substitution
swap_color	103	visual swap (tests binding)
swap_material	101	visual swap (tests binding)
swap_spatial	100	spatial swap
swap_role	45	role swap (multi-agent scenes)

Note on swap_role (n=45): role swaps require ≥2 distinct agents in the scene and are therefore structurally constrained by the fraction of multi-subject scenes in the benchmark. This is a dataset property, not a generation shortfall. Results on swap_role are reported in the appendix of the accompanying paper with a disclosure footnote; main-paper tables use n ≥ 50 as the cutoff.

Example usage

import pandas as pd
from PIL import Image
from datasets import load_dataset

# Via huggingface datasets
ds = load_dataset("perona-lab/CompoSET", split="test")
row = ds[0]
img_base = row["image_base"]           # auto-decoded PIL.Image (if using Image() feature)
captions_medium_var = row["captions"]["medium"]["var"]

# Or direct parquet read (paths are strings, resolve yourself)
df = pd.read_parquet("variations.parquet")
for r in df.itertuples():
    base = Image.open(r.image_base)
    var  = Image.open(r.image_var)
    for tier in ("short", "medium", "long"):
        c_base = r.captions[tier]["base"]
        c_var  = r.captions[tier]["var"]
        # evaluate here — supports caption-pair, image-pair, or 2x2 (Winoground) scoring

Source data and generation

Each scene was generated via a Claude + Gemini pipeline with three QC gates. A full corrected scene caption (shipped in scene_captions.parquet) drove image generation; variations apply a single localized edit via prompt modification; generated pairs went through manual QC for cardinality, drift, edit-match and disambiguation. Discarded variations are not shipped in this release. Overall discard rate: 22.4% (513 discarded / 2,289 attempted).

Method details (full pipeline, prompts, QC rubric) are documented in the accompanying paper.

Splits

Single test split. CompoSET is intended purely as an evaluation benchmark; there is no training split. Do not fine-tune on this data.

Intended uses

• Evaluate compositional binding / attribute-object binding / spatial grounding in vision-language models.
• Winoground-style scoring (I2T / T2I / Group, no partial credit).
• Derived views: caption-pair (ARO-style), image-pair (BiVLC/SugarCrepe++-style), 2×2 (Winoground).
• Per-edit-type failure analysis and per-scene difficulty disambiguation (made possible by the many-foils-per-scene property).

Limitations

• swap_role is under-sampled (n=45) by dataset structure (multi-agent scenes are a subset of total scenes). Appendix-only in the paper.
• Generated images. Base and variation images are produced by a state-of-the-art image generator (gemini-2.5-flash-image, known as "nanobanana"). A drift audit of N=300 sampled released pairs found 2 pairs (0.67%, Wilson 95% CI [0.18%, 2.39%]) with incidental image changes beyond the intended single edit. Disocclusion artifacts (e.g. a detail becoming visible because an occluder moved) are excluded by rubric and not counted as drift. Details in the paper.
• English only. All captions are in English; multilingual extension is future work.
• No training split. Do not train on this data. Using CompoSET as training data would defeat its evaluation purpose and risk contamination of future reported numbers.
• Single-edit by construction. CompoSET deliberately isolates one compositional axis at a time. It does not probe multi-edit compound compositionality; that is a separate benchmark design question.

Responsible AI considerations

• No personally identifying information. All persons depicted are generated (synthetic), not real.
• Gendered terminology banned. Captions use gender-neutral terms ("person", "the person") throughout; "man" / "woman" / "boy" / "girl" are not used. This is both a bias-mitigation choice and a caption-consistency choice.
• No sensitive categories (age, gender, emotion, political, religious, medical) are used as edit attributes. The taxonomy explicitly excludes age / gender / emotion.
• Synthetic image generation leverages a commercial API (Gemini 2.5 Flash Image). Users must comply with Google's terms of service when redistributing or using these images. Images are released under CC-BY-4.0 consistent with the underlying API terms for generated content.
• Generator drift is measured and disclosed (6.33% at short-tier captions; see Limitations).

License

CC-BY-4.0 for captions, metadata, and per-scene descriptions. Generated images are released under the same terms, consistent with the image generator's terms of service at the time of generation.

Citation

Paper and BibTeX forthcoming — submission target NeurIPS 2026 Datasets & Benchmarks Track. Please cite as:

@misc{composet2026,
  title  = {{CompoSET}: Compositional Single-Edit Testbed for Vision-Language Models},
  author = {Laure Delisle and Pietro Perona},
  year   = {2026},
  note   = {(soon to be under review)}
}

Version

v1.0 (2026-04-23) — 80 scenes, 1,776 variations, 16 edit types, short/medium/long captions.