Datasets:

zhuxiangru
/

SemVarBench

license: cc-by-4.0
task_categories:
  - text-to-image
  - text-generation
language:
  - en
tags:
  - compositionality
  - semantic-variation
  - text-to-image
  - benchmark
  - vision-language
pretty_name: SemVarBench
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files:
      - split: train
        path: train.parquet
      - split: test
        path: test.parquet

SemVarBench

SemVarBench is the benchmark from the ICLR 2025 paper Evaluating Semantic Variation in Text-to-Image Synthesis: A Causal Perspective, designed together with the SemVarEffect metric to evaluate the causality between semantic variations in the input text and the generated image in text-to-image (T2I) synthesis.

Each example is built around a base caption T0 and a minimally-edited variant T1 that changes the composition (e.g. swapped subject/object or swapped attributes) while reusing the same words, plus T2, a paraphrase of T1 (passive voice / reordering) that is semantically equivalent to T1. Semantic variations are achieved through two types of linguistic permutations while avoiding easily predictable literal variations.

This dataset is the flattened version of the benchmark/ directory in the SemVarBench GitHub repository, merging:

trainingset/training_data.txt — the train split (10,770 rows).
testset/test_data.txt — the test split (684 rows).
testset_divided_category/ — 20 per-category slices of the test set, used here to populate the categories field.

Data fields

Column	Description
`id`	Unique example identifier (e.g. `0_61_326`).
`T0`	Base caption.
`T1`	Semantically varied caption (minimal compositional edit of `T0`).
`T2`	Paraphrase of `T1` (passive / reordered), semantically equivalent to `T1`.
`categories`	List of contrast-category tags. Populated for the test split; empty (`[]`) for the train split. A test example may carry more than one tag (118 of 684 do).

The 20 test categories

absolute_location, action, age, appearance, color, counting, direction, height, interaction, manner, material, relative_location, sentiment, shape, size, spatio_temporal, temperature, texture, vague_amount, weight.

Splits

Split	Rows
`train`	10,770
`test`	684

Train and test ids are disjoint. The 20 category files together cover exactly the 684 test ids (no more, no fewer).

Usage

from datasets import load_dataset

ds = load_dataset("zhuxiangru/SemVarBench")
print(ds["test"][0]["T0"], "||", ds["test"][0]["T1"], "||", ds["test"][0]["T2"])

# filter the test set to a single category
color = ds["test"].filter(lambda r: "color" in r["categories"])
print(len(color), "color examples")

Related work

The predecessor dataset Winoground-T2I is also available on the Hub at zhuxiangru/Winoground-T2I.

Citation

If you find the data in our project useful, please consider citing our work:

@inproceedings{DBLP:conf/iclr/ZhuSSXL00YX25,
  author       = {Xiangru Zhu and
                  Penglei Sun and
                  Yaoxian Song and
                  Yanghua Xiao and
                  Zhixu Li and
                  Chengyu Wang and
                  Jun Huang and
                  Bei Yang and
                  Xiaoxiao Xu},
  title        = {Evaluating Semantic Variation in Text-to-Image Synthesis: {A} Causal
                  Perspective},
  booktitle    = {The Thirteenth International Conference on Learning Representations,
                  {ICLR} 2025, Singapore, April 24-28, 2025},
  publisher    = {OpenReview.net},
  year         = {2025},
  url          = {https://openreview.net/forum?id=NWb128pSCb},
  timestamp    = {Thu, 15 May 2025 17:19:05 +0200},
  biburl       = {https://dblp.org/rec/conf/iclr/ZhuSSXL00YX25.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}