FlowDIS-DIS5K

DIS5K images and masks joined to the language prompts of PAIR/FlowDIS-DIS5K-language-prompts, so language-guided dichotomous segmentation needs one load_dataset call instead of a join. 5470 rows, all six DIS5K splits.

from datasets import load_dataset

ds = load_dataset("nobg/FlowDIS-DIS5K", split="DIS_VD")
ds[0]["image"]            # PIL, original resolution
ds[0]["label"]            # PIL, the binary mask
ds[0]["prompt"]           # 'Four white wind turbines with red markings ...'
ds[0]["prompt_keywords"]  # the short form -- see below

Why this exists: the upstream test split collapses four splits into one

HuggingFace infers splits from file names, and FlowDIS ships one JSON per DIS5K split. Four of the six names collapse:

upstream file	inferred split	rows
DIS-TR.json	train	3000
DIS-VD.json	validation	470
DIS-TE1.json … DIS-TE4.json	test — all four merged	2000

So load_dataset("PAIR/FlowDIS-DIS5K-language-prompts")["test"] is a 2000-row mixture of four splits of graded difficulty (DIS-TE1 easiest → DIS-TE4 hardest) — precisely the axis DIS5K's test splits exist to separate. It reports one number where the benchmark defines four, and nothing warns you.

This dataset keys off each row's own split field instead, and uses nobg/DIS5K's six split names, so it lines up 1:1 with the image source.

The join is exact

filename (FlowDIS) matches image_name (DIS5K) byte-for-byte, taxonomy prefix included — 22#Weapon#3#Missile#11711594126_3831657d91_o.jpg. Measured across all six splits: 5470 / 5470 matched, 0 unmatched, 0 duplicates. No normalization or fuzzy matching. The build asserts the tally per shard and refuses to write a partially-joined split, so an empty prompt cannot occur.

Image and label bytes are copied through undecoded — pixels are bit-identical to nobg/DIS5K, with no re-encode or resize.

The three prompt columns

Phrasings per row differ by split, and only DIS_TR has more than one:

split	rows	phrasings/row
DIS_TR	3000	4
DIS_VD	470	1
DIS_TE1–DIS_TE4	500 each	1

column	type	contents
prompts	list<string>	all phrasings, verbatim — lossless
prompt	string	prompts[0], the full descriptive sentence
prompt_keywords	string	prompts[1] where it exists, else prompts[0]

In DIS_TR, index 0 is a descriptive sentence, index 1 a comma-separated keyword form, and 2–3 paraphrases of 0:

prompt          'A silver missile mounted on a large green military transport vehicle.'
prompt_keywords 'Missile, military launcher vehicle'

prompt_keywords is not cosmetic. SAM 3's processor tokenizes at padding="max_length", max_length=32, so descriptive sentences truncate silently while the keyword form fits. Note the trade-off in its fallback: on the five eval splits (1 phrasing each) it is identical to prompt. Stated plainly because a column that is sometimes a copy is otherwise a trap.

flowdis_split records the upstream label each row declared, so the collapse above stays auditable from inside the data.

Licensing — read before redistributing

The upstream prompt repo declares no license. The prompts are redistributed here on that basis, with attribution below; images and masks come from nobg/DIS5K (Apache-2.0, itself subject to DIS5K's own Terms of Use). Treat the prompts as a soft dependency: nobg's training script keeps --flowdis off by default and falls back to a concept derived from the filename taxonomy, and anything built on this should preserve that property.

Citation

Prompts — FlowDIS:

@article{sargsyan2026flowdis,
  title={{FlowDIS: Language-Guided Dichotomous Image Segmentation with Flow Matching}},
  author={Sargsyan, Andranik and Navasardyan, Shant},
  journal={arXiv preprint arXiv:2605.05077},
  year={2026},
  url={https://arxiv.org/abs/2605.05077}
}

Images and masks — DIS5K:

@inproceedings{qin2022dis,
  title={Highly Accurate Dichotomous Image Segmentation},
  author={Qin, Xuebin and Dai, Hang and Hu, Xiaobin and Fan, Deng-Ping and Shao, Ling and Van Gool, Luc},
  booktitle={ECCV},
  year={2022}
}