neulab/t-dig-test

T-DIG interactive test set

Prompts and diversity requests written by people using an image generator, collected for T-DIG: Text-conditioned Diversity for Image Generation (under review).

Each item is one prompt a participant typed, the four images the T2I model produced for it by independent sampling, and the participant's request for what should differ across the images (the axis), in their own words. These are the prompts and axes of the paper's test-set results and human evaluation.

T2I model	Items	Participants	First request was "generate"	Later requests
Qwen-Image-2512	105	21	75	29
Stable Diffusion 3.5 Large	101	20	62	52
Stable Diffusion 3.5 Medium	99	7	42	35
FLUX.2 [klein] 9B	101	24	71	53

The paper's automatic test-set metrics use the items whose first request was generate.

How it was collected

Participants recruited on Prolific used an interactive tool backed by one T2I model. They typed a prompt, saw four i.i.d. samples, and wrote in their own words what should vary across the set. They could then either generate a new set (generate) or edit the four images on screen (edit), and keep refining. Prompts and axes are verbatim, including spelling.

Fields

Field	Description
item_id	<T2I model>-<number>. Numbers have gaps where items were withheld (see Safety).
renderer	The T2I model that drew iid_images.
prompt	The prompt, as typed.
axis	The participant's first request for what should differ, as typed.
mode	generate (a new set from the prompt) or edit (edit the four images) for that first request.
later_requests	Further requests on the same prompt, in order: axis, mode, and applied_to (original images or previous round).
participant	Anonymous label (p01, p02, …), assigned per T2I model; labels are not linked across T2I models.
iid_images	Paths of the four i.i.d. samples (1024×1024 PNG) the participant saw, relative to the T2I model's folder.

from datasets import load_dataset
ds = load_dataset("neulab/t-dig-test", "qwen-image-2512", split="test")

Safety and privacy

• Participants: no Prolific IDs, session IDs or other identifiers are included, only per-T2I-model labels.
• Prompt text: scanned for e-mail addresses, links, phone numbers, handles and named public figures; none remain.
• Images: every image was screened with a vision-language classifier for nudity, minors and violence, and the flags were reviewed by hand.
• Withheld items: four items were removed and their numbers left as gaps:
  • one depicting a named real person
  • three depicting minors or young-looking people in underwear or swimwear
• Remaining content: some items show mild fantasy or action violence (battle scenes, monsters) or adults in swimwear or revealing clothing. They were reviewed and kept.

Licenses

• Text: prompts and axes are released under the license above.
• Images: generated by Qwen-Image-2512, Stable Diffusion 3.5 Large and Medium, and FLUX.2 [klein]; their use is also subject to each model's license.

Citation

Citation information will be added after review.