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.

This dataset contains the stimuli for a human study that measures it. Each stimulus is an (image, prompt) pair labeled with a case from a taxonomy of unmentioned content, and most stimuli belong to a minimal pair: the same prompt with a test image and a reference image, or the same image under two prompts. The quantity of interest is how human ratings move between the two members of a pair.

Status: pilot. The current images are placeholder SVG illustrations rendered to PNG, used to validate the protocol. They will be replaced by real generations and inpainted minimal pairs. Annotations are not included yet.

Taxonomy

Category	Name	Definition	Cases
A	Forced completion	The prompt leaves a dimension open and the image has to fill it with a typical value	A1–A3
A′	Atypical completion	The same forced dimension, filled with a valid but atypical value	A′1–A′3
B	Implied content	Content the prompt never names but entails or excludes	B1–B3
C	Optional addition	Content that nothing in the prompt requires	C1–C9
D	Salience	Nothing is added or removed; what changes is what the image is about	D1–D2
E	Control	Attention check	E1
G	Generator-specific	Extra content typical of diffusion outputs	G1–G4

Contents

data/
  metadata.jsonl       one line per stimulus (the source of truth)
  images/              image files referenced by `file_name`

data/ is a standard imagefolder: the Hub converts it to Parquet for the viewer and for load_dataset, so there is no build step.

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–S44; S01–S12 are the pilot samples
prompt	str	Text shown to the annotator
case_id	str	Taxonomy case, e.g. C2
category / category_name	str	A, A′, B, C, D, E, G and its name
subcategory	str	Case name
visual_style	str	Visual style of the image in one word (svg-illustration for now)
role	str	test, minimal_pair (reference) or control
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
image_source	str	placeholder_svg_illustration for now

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"])

Citation

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