Enzo-py/partial-prompting-ds

Partial Prompt Human Study

How do humans judge content that a text-to-image prompt does not mention?

Prompts given to diffusion models are almost always partial: "a chair on the grass" says nothing about the sky, and "an apple" says nothing about whether a pear may sit next to it. Every generated image has to fill these gaps. Existing T2I alignment benchmarks (TIFA, DSG, Gecko, VQAScore, RichHF) derive their questions from the prompt alone, so unmentioned content is invisible to them by construction, and human annotation guidelines settle the question by convention instead of measuring it.

Taxonomy

Each stimulus is described along three dimensions, which are the axes of analysis:

1. Relation (relation): how the image stands to the prompt.
   • match: every stated element is shown, and the unmentioned content is only a plain or simple setting. "an apple" → an apple on white; "a polar bear" → on sea ice.
   • addition: the image shows more than the prompt; the extra content is true, unmentioned and notable. "a laptop on a desk" → a lamp and a mug too.
   • mismatch: the image does not do what the prompt says. Its mismatch_type keeps the detail: absent, a named or entailed element is missing ("an apple and a mug on a table" → no mug), or wrong, the image states the opposite ("a cat to the left of a dog" → the cat on the right; "with no lamp" → a lamp; "a chair" → two chairs).
2. Target (target): what the gap is about: background, object, count, attribute or relation (null for match), grouped in target_group: setting (background), content (object, count), property (attribute, relation).
3. Modulators (modulators): what makes the gap more or less visible: typicality: atypical (an unusual or incongruous completion), salience: low (the named subject is small or secondary), prompt_detail (partial, detailed), constraint: negation ("lonely", "with no …").

The omission quads compare addition with an absent mismatch directly (see Controls below). The finer cases below are kept in case_id for compatibility; tools/label_dimensions.py derives the dimensions from them.

Each stimulus is also labeled with one case (case_id) and, for cases that have them, a variant (case_variant): 12 cases in 6 categories. Examples read prompt → what the image shows.

• A — Forced completion: the prompt leaves a dimension open and the image has to fill it, with a typical or atypical value.
  • A1 Unspecified attribute or kind: the image picks a value the prompt leaves open: a property, or which member of the named category. "an apple" → a green apple (typical); "a bird" → a penguin, "an apple" → a blue apple (atypical). Variants: typical, atypical.
  • A2 Background / framing: some background, angle and scale have to be chosen. "a chair on the grass" → sky, sun, cloud, tree (typical); "a skyscraper" → alone in a desert (atypical); "a cup of coffee" → seen from directly above (viewpoint). Variants: typical, atypical, viewpoint.
• B — Implied content: content the prompt never names but entails or excludes.
  • B1 Implied presence: content the prompt entails is shown, or is missing. "a birthday cake" → lit candles (realized) or no candles (missing). Variants: realized, missing.
  • B2 Absence, violated: the prompt excludes content that appears anyway, through a word (implicit) or a negation (explicit). "a lonely tree in a field" → a second tree (implicit); "a laptop on a desk, with no lamp" → a desk lamp (explicit). Variants: implicit, explicit.
• C — Optional addition: content that nothing in the prompt requires.
  • C1 Added object: an object that is typical of the scene, out of place (semantic) or in an impossible position (syntactic); animate additions: see the Animacy axis. "a laptop on a desk" → a desk lamp and a mug (typical); "an apple on a table" → a traffic cone (incongruous-semantic) or a mug floating above the table (incongruous-syntactic). Variants: typical, incongruous-semantic, incongruous-syntactic.
  • C2 Competing object: same level and salience as the named subject. "an apple" → an apple and a pear.
  • C3 Extra instance: more than one of the named entity; with a stated relation, the extra instance can make it ambiguous. "a chair on the grass" → two chairs; "a cat to the left of a dog" → two dogs, the cat between them (relation). Variants: plain, relation.
  • C4 Added relation: the named objects are right; the relation is invented. "a cat and a box" → the cat inside the box.
  • C5 Modified subject: an accessory or a state change on the named entity itself. "a dog" → a dog wearing sunglasses.
• D — Salience: nothing is added or removed; what changes is what the image is about.
  • D1 Subject salience: the named subject is visually minor. "a sailboat" → a tiny boat on the horizon (small); "a handbag" → a woman holding the bag dominates the image (inverted). Variants: small, inverted.
• F — Completeness vs correctness: the same image under prompts that trade completeness against truth.
  • F1 Partial but true vs detailed but wrong: a partial prompt whose details are all true (reference) against prompts that state x more details, exactly one of them wrong (test), and, in three series (S140, S151, S167), the same detailed prompt with every detail true (control), to separate the effect of the error from the effect of a longer prompt. "a laptop on a desk" vs "a laptop, a desk lamp, a mug and a vase on a desk" → a laptop, a lamp and a mug, no vase. Is a wrong detail worse than missing ones? Variants (the kind of error): missing (a named object is absent), attribute, count, relation.
• G — Generator-specific: content typical of diffusion outputs.
  • G1 Attribute spill-over: an explicit attribute bleeds into unmentioned content. "a red apple" → a red table and a red wall.

Controls are not cases: their prompt is plainly contradicted by the image (role = control, case_id = null, the error in control_type: absent, relation, attribute, binding or count). "a banana" → an apple; "a cat to the left of a dog" → the cat on the right; "two apples" → three apples. Their answer is known in advance: they are attention checks for annotators and check that a metric drops; 5 of them are the "missing from the image" cell of an omission quad: the same pair of objects (A, B) under the prompts "A" and "A and B", on an image of A and an image of A and B, so a missing object can be compared with an unmentioned one (factors.quad, factors.omission).

The category letters are kept from earlier versions (there is no E), so that pair ids such as P-E1 or P-C6 stay stable.

Cross-cutting axes vary one factor across cases; their pairs have X-* ids.

• Prompt specificity: "an apple" vs "a red apple on a wooden table", on the same image. Are detailed prompts read as exhaustive?
• Exhaustivity: "apple" · "an apple" · "a single apple" · "an apple among other fruit". If cancelling the implicature removes the penalty, the effect is pragmatic rather than semantic.
• Prompt type: an object prompt ("a mug") vs a scene prompt ("a kitchen"). How much unmentioned content does a scene license?
• Animacy: an inanimate vs an animate addition of matched size. Are animate additions penalized more?
• Dose: 0 · 1 · 3 · 10 added objects, or a small vs a large area. A penalty curve instead of a binary effect.
• Reference zero: the subject alone on a neutral background, a floor for measuring the effect of any completion.

Statistics

179 stimuli on 103 images (104 photo, 75 SVG). tools/stats.py draws these figures and more (score distributions, correlations) into stats/.

Relations and targets. The three relations are of similar size: 59 match, 58 addition, 62 mismatch (17 absent, 45 wrong). Additions are mostly about the background and added objects; mismatches cover every target.

Modulators. Each modulator value holds 12 to 16 stimuli; most stimuli carry none.

SVG vs photo. Both styles cover the three relations and the five scores, with more mismatches among photos: compare styles within a relation, not overall.

Image reuse. An image serves at most 4 stimuli (the same image under different prompts, or the members of a prompt pair); most images serve one or two.

Contents

data/
  metadata.jsonl       one line per stimulus (the source of truth)
  images/              image files referenced by `file_name`
  train-00000-of-00001.parquet   the same rows with the images embedded (config `default`)
pairs/
  pairs.jsonl          one line per minimal pair (built from metadata.jsonl)
groups/
  groups.jsonl         one line per group: 3 to 5 images to rank under one prompt, or 3 to 5 prompts on one image
segmentation/
  <image_id>.png       segmentation of each image (COCO panoptic: segment id per pixel)
  <image_id>.json      its segments (COCO panoptic segments_info)
  categories.json      the 75 categories
figures/               the figures of this README (tools/stats.py)

data/metadata.jsonl and data/images/ are the source of truth (a standard imagefolder, used by the scripts). The default config reads data/train-00000-of-00001.parquet, built from them with the image bytes embedded, so the viewer and load_dataset do not fetch the PNG files one by one.

Fields

Field	Type	Description
image	Image	The stimulus. In data/metadata.jsonl this column is file_name, a path relative to data/
sample_id	str	S01–S227, with gaps (removed rows keep their id free); S01–S12 are the pilot samples
prompt	str	Text shown to the annotator
case_id	str | null	Taxonomy case (12 cases, listed above), e.g. C2; null for a control
category / category_name	str | null	A, B, C, D, F, G and its name; null for a control
subcategory	str | null	Case name; null for a control
case_variant	str | null	Variant of the case: typical / atypical (A1), typical / atypical / viewpoint (A2), realized / missing (B1), implicit / explicit (B2), typical / incongruous-semantic / incongruous-syntactic (C1), plain / relation (C3), small / inverted (D1), missing / attribute / count / relation (F1, the kind of error)
relation	str	match, addition or mismatch (see Taxonomy)
mismatch_type	str | null	For a mismatch: absent (a named element is missing) or wrong (the image states the opposite); null otherwise
target	str | null	background, object, count, attribute or relation; null for match
target_group	str | null	setting (background), content (object, count) or property (attribute, relation); null for match
modulators	dict | null	Any of typicality (atypical), salience (low), prompt_detail (partial / detailed), constraint (negation)
hypothetical_score	int	The rating (1–5) expected before the study, from tools/hypothetical_score.py: 5 aligned, setting only; 4 notable unmentioned content (not extra instances of a singular subject: a count error, 2), a small or secondary subject, or an atypical completion; 3 one error among several correct named objects, an ambiguous relation, or an incongruous addition; 2 one named object right but its attribute, relation, count or setting wrong, or one of two named objects absent; 1 nothing named shown, or a violated negation or exclusivity (lonely, single, with no other …) on the only named object
visual_style	str	Visual style of the image in one word: svg-illustration (drawn, image_source placeholder_svg_illustration) or photo (generated, generated:<model>)
role	str	test, minimal_pair (reference) or control
control_type	str | null	For a control (prompt plainly contradicted): absent, relation, attribute, binding or count; null otherwise
attention_check	bool	An attention check: a control whose answer is obvious, outside the omission quads (20). The lists are built from the other rows (study items, one place each) plus two attention checks drawn at random per list
list	int	The annotation list (0–9) the stimulus belongs to, from tools/assign_lists.py: 18 stimuli per list (17 in the last), 2 of them attention checks; inside a list no image appears twice and no two members of a pair meet
pair_ids	list[str]	Minimal pairs the stimulus belongs to; P-* pairs are defined by a case, X-* by a cross-cutting axis
pair_axes	list[str]	For each pair, what varies: image or prompt
pair_roles	list[str]	For each pair, this stimulus's role in it: test, reference or control (a stimulus can be the test in one pair and the reference in another)
unmentioned_content	str	What the image contains that the prompt does not state
image_id	str	Shared by stimuli that reuse the same image under different prompts
factors	dict | null	Levels of the factorial series: relation (correct / inverted), binding (correct / swapped), count (correct / wrong), extra (unmentioned object), background (clean / simple / rich), added (F1: number of details the prompt states beyond the partial reference, 0 for the reference) error (F1: kind of the wrong detail, null when every detail is true), and quad / omission (omission quads: the prompt names A or A and B, the image shows A or A and B; omission is none when both match, prompt when the prompt leaves B out, image when the image leaves B out)
image_source	str	placeholder_svg_illustration (drawn from SVG, segmented exactly) or generated:<model> (text-to-image, e.g. generated:qwen2.1)
description	str | null	Long human description of a generated image
generation_prompt / generation_seed	str / int | null	Prompt and seed used to generate the image; the prompt shown to annotators is prompt
segmentation	str	Segmentation of the image: segmentation/<image_id>.png, with its segments in the .json of the same name (path relative to the repository root)
objects	list	Every object in the image: segment_id (its id in the segmentation), label, mentioned (named in the prompt), negated (named under a negation, e.g. "with no lamp"), bbox ([x, y, w, h] in pixels), area (pixels), attributes (JSON string or null, e.g. {"depicted": true}, {"state": "wrecked"})

Pairs

Two background series keep a stated relation fixed: "a cat to the left of a dog" (clean, rich) and "an apple to the left of a mug" (clean, simple, rich). Whether a metric ignores the background is tested by X-BG-*; whether it catches a wrong relation, by X-REL-* ("a cat next to a box" on a cat in a box, the lantern on a rock) and P-E2b. Binding (X-BIND-*) and counting (X-COUNT-*) pairs follow the same idea; X-COUNT-PEAR asks whether an unmentioned pear changes "two apples".

pairs/pairs.jsonl (config pairs) has one line per minimal pair:

Field	Type	Description
pair_id	str	P-*: defined by a taxonomy case; X-*: by a cross-cutting axis (X-SPEC* specificity, X-ORDER-* order of mention, X-REL-* stated relation correct vs inverted, X-BIND-* colours bound correctly vs swapped, X-COUNT-* counting, X-OMIT-IMG-* / X-OMIT-PROMPT-* an object missing from the image vs missing from the prompt (see factors.omission), X-ATTR-* / X-STATE-* / X-CTX-* stated attribute, state or setting right vs wrong, X-EXH-* exhaustivity ("a single apple"), X-LONELY-* "lonely", X-NEG-* explicit negation respected vs violated, X-NEGBG-* the negated content only in the background, X-NEGTYPE-* explicit ("with no other buildings") vs implicit ("lonely") negation, X-TYPE-* object vs scene prompt, X-BG-* the same scene on a clean, simple and rich background, X-DOSE-*, X-ANIM-*)
axis	str	image (same prompt, two images) or prompt (same image, two prompts)
family_expected	str	invariance (the varied content is licensed by the prompt: a fair metric should not move), sensitivity (the prompt is contradicted: it should drop) or open (humans may or may not penalize the unmentioned content; this is what the study measures)
family_human	str | null	The same, as decided by the human ratings (filled after the study)
cases	list[str]	Taxonomy cases of the members (control for a control)
prompt / image_id	str	What is shared: the prompt of an image pair, the image of a prompt pair
varies	str	What differs between the members
controlled	bool	true when the members differ only by what the pair varies (same SVG drawing, or the same image under two prompts); false for independent generations that share a prompt
members	list	sample_id and role (test, reference, control)

Groups

A second task: the annotator ranks 3 to 5 images from best to worst for one prompt. Each group keeps the prompt fixed and varies what the prompt leaves open (which bird, how salient the subject is, what surrounds it), so the ranking shows which completions people prefer, not only which ones they accept. A prompt group does the reverse: one image, 3 to 5 prompts to rank by how well each describes it (GR-WOMAN-CHAIR-PROMPTS: is a short true prompt ranked above a detailed prompt with one wrong detail?). groups/groups.jsonl (config groups) has one line per group:

Field	Type	Description
group_id	str	GR-*
axis	str	image (rank images under one prompt) or prompt (rank prompts on one image)
prompt	str | null	The prompt shown above the images; null for a prompt group
visual_style	str	svg-illustration or photo: a group never mixes the two
varies	str	What differs between the images
ranking_human	list[str] | null	The images (or, for a prompt group, the prompts) in the order the annotators chose (filled after the study)
ranking_hypothetical	list[str]	The order expected before the study (tools/hypothetical_score.py): mostly by hypothetical_score, with a simple setting ranked above a rich one, and a rich one above a plain or empty one
members	list	image_id, file_name (path relative to data/), prompt and sample_id: the stimulus with this image and this prompt when there is one, else null (rated only in the group)

Segmentation

Every image is segmented in the COCO panoptic format, one pair of files per image_id in segmentation/:

• <image_id>.png: each pixel holds the id of the object it belongs to, id = R + 256·G + 256²·B (0 = nothing);

• <image_id>.json: segments_info, one entry per object (and, for photos, source: the segmenter): id, category_id, category, isthing (0 for background regions such as sky, wall, floor or grass), bbox ([x, y, w, h] in pixels), area, attributes (e.g. {"depicted": true}, {"state": "wrecked"});

• categories.json: the category list (id, name, isthing).

• SVG images (visual_style = svg-illustration): the masks come from the same drawing as the image, so they are exact and align to the pixel; cast shadows belong to no object.

• Photos (visual_style = photo): the masks are automatic (Mask2Former Swin-L, COCO panoptic), imported by tools/import_panoptic.py, and not hand-checked. Labels are mapped to the same vocabulary (couch → sofa, wall-other → wall) but stay at COCO's granularity (a flamingo is a bird, a beer glass a cup, a skyscraper a building), some objects are missed (the shutters of G15 / G18), and some pixels belong to no segment (54% of G34). Stimuli that share an image share its segmentation: each row of metadata.jsonl points to it (segmentation) and lists its objects (objects) with mentioned / negated computed from that row's prompt.

import json, numpy as np
from PIL import Image
row = ...                                        # a row of data/metadata.jsonl, e.g. S46
seg = np.array(Image.open(row["segmentation"]), dtype=np.uint32)
ids = seg[..., 0] + 256 * seg[..., 1] + 256 ** 2 * seg[..., 2]
masks = {o["label"]: ids == o["segment_id"] for o in row["objects"]}
unmentioned = [o["label"] for o in row["objects"] if not o["mentioned"]]

Usage

from datasets import load_dataset
ds = load_dataset("Enzo-py/partial-prompting-ds", split="train")
pairs = ds.filter(lambda x: "P-C2" in x["pair_ids"])
table = load_dataset("Enzo-py/partial-prompting-ds", "pairs", split="train")
invariance = table.filter(lambda p: p["family_expected"] == "invariance")

To look at a stimulus or a pair (image, prompt, segmentation, with the objects named in the prompt in green and the unmentioned ones in orange), from a local copy of the repository:

pip install matplotlib pillow numpy
python scripts/visualize.py S46                       # one stimulus
python scripts/visualize.py X-BG-CAT-LEFT-DOG         # all members of a pair
python scripts/visualize.py --list                    # every stimulus and pair
python scripts/visualize.py S46 --out s46.png         # save instead of showing

Citation

@misc{partialprompt2026,
  title  = {Partial Prompt Human Study},
  author = {TODO},
  year   = {2026},
  note   = {Pilot release}
}

Score hypotheses

Before the study, every stimulus gets the rating we expect from annotators (hypothetical_score, 1 to 5) and every group the order we expect (ranking_hypothetical), both written by tools/hypothetical_score.py. The study then confirms or refutes them; they are hypotheses, not labels.

Score	Hypothesis	Examples
5	The image matches the prompt; the unmentioned content is only a plain or simple setting	"an apple" → an apple on white; "a polar bear" → on sea ice
4	Notable unmentioned content (a rich background, extra objects), a small or secondary subject, or an atypical completion	"a laptop on a desk" → a lamp and a mug too; "a bird" → a penguin; "a handbag" → small in a corridor
3	One error among several correct named objects independent of it, an ambiguous relation, or an incongruous addition	a detailed prompt with one wrong detail; "a blue mug and a red bowl" → colours swapped; "a cat to the left of a dog" → a cat between two dogs
2	One named object is right but its attribute, relation, count or setting is wrong, or one of two named objects is absent	"a blue apple" → a red apple; "a cat in a box" → next to the box; "a skyscraper in a city" → in a desert; "a chair" → two chairs; "a cat and a box" → no box
1	Nothing named is shown, or a negation or exclusivity on the only named object is violated	"a banana" → an apple; "a lonely tree" → two trees (the evaluated object is "a tree with no other tree", which nothing matches)

Rules behind the table:

• Counting. Several instances of a singular subject are a count error, not an addition (score 2).
• Negation. An implicit negation (lonely, single) inherits the rule of an explicit one (with no …): what is evaluated is the negated object, so a violation on the only named object leaves nothing aligned (score 1). Next to other correct named objects ("a laptop on a desk, with no lamp"), it follows the error rule (score 3).
• Background. Rated: a simple background 5, a plain one 5, a rich one 4. Ranked: simple, then rich, then plain: annotators are expected to prefer some context to none, while still rating a busy background a little lower.
• Detail. Among true prompts on the same image, the most detailed is expected to rank first; a detailed prompt with one wrong detail is expected below every true prompt, however short (GR-WOMAN-CHAIR-PROMPTS).

Distribution: 9 stimuli at 1, 35 at 2, 22 at 3, 54 at 4, 59 at 5 (mean 3.66). A constant prediction of 4 has a mean absolute error of 0.99 and a constant 5 of 1.34: a useful metric has to beat both.

Some pairs labelled invariance (a richer background or an added object, P-A2, P-C1, X-BG-*) expect a one-point drop under these hypotheses: the pair table assumes that a fair metric ignores the background, the hypotheses that annotators do not quite. The study decides.