Hugging Face's logo

Datasets:
Gemus-AI
/
DesignAnything-Bench

Modalities:
Image
Text
Formats:
json
Languages:
English
Size:
< 1K
ArXiv:
Tags:
design
benchmark
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
DesignAnything-Bench / README.md
MizuharaShigen
Link the Design Anything paper (arXiv 2610.24771)
674a8e7 verified
preview code
|
Raw History Blame Contribute Delete
3.93 kB
metadata 
license: other
license_name: per-file
license_link: >-
  https://huggingface.co/datasets/Gemus-AI/DesignAnything-Bench/blob/main/LICENSES.md
pretty_name: Design Anything Bench Inputs
language:
  - en
tags:
  - design
  - benchmark
size_categories:
  - n<1K
configs:
  - config_name: images
    default: true
    data_files:
      - split: inputs
        path: metadata/images.jsonl

Design Anything Bench Inputs

The input images that the 228 tasks of the Design Anything benchmark (corpus v0.6) provide to the system under test: photographs, drawings, maps, references and generated stand-ins that the briefs ask the designer to work from.

Related material:

Contents

Content Files Size
Input images (blobs/) 682 1.16 GB 
metadata/images.jsonl                      one row per image (the dataset viewer shows these)
blobs/sha256/<first two hex digits>/<sha256>.<ext>   the images, byte-identical
LICENSES.md                                licenses and third-party attribution
CHECKSUMS.sha256                           SHA-256 of every other file

Each task's brief (benchmark/tasks/<taskId>/scenario.json in the GitHub repository) lists the images it provides in referenceAssets, with a path and a sha256. The image with that digest is blobs/sha256/<first two hex digits>/<sha256>.<ext> here. The images are byte-identical to what the systems received, so metadata embedded in them (for example EXIF) is kept as well.

Row fields (metadata/images.jsonl)

  • sha256, path, bytes, contentType.
  • tasks: the tasks that provide the image; corpusPaths: the referenceAssets[].path values that name it.
  • origin: authored (made by the benchmark authors), generated (generated by the benchmark authors with an image model), or third-party.
  • license, source, licenseUrl, attribution: see below.
  • redistributionReview: reviewed for every image in this release (see below).
  • flags: share-alike, url-source (the record points to a source web page), label-conflict (two records give different license labels), exif / xmp (the file carries embedded metadata).

Licenses

All 682 inputs were reviewed for redistribution on 2026-10-17. The authors confirmed the authored and generated images. Each of the 55 third-party images was checked against its source page: the license on the page matches the record, and the CC BY-SA version and author match the attribution. The aerial photograph site-overhead.png for the architecture base-A tasks is USDA NAIP imagery, a US federal government work; it was labeled CC0 in the task corpus and is listed here as public domain.

Seven inputs carry CC BY-SA and two carry OGL-Toronto-1.0; their original attributions are listed separately in LICENSES.md. The third-party images that carry EXIF or XMP metadata are byte-identical to the files published at their sources.

This dataset is not released under a single license. Every image keeps its own license, recorded per file:

  • authored and generated images: CC BY 4.0, attribution "The Design Anything Authors".
  • third-party images: the license of their source (public domain, CC0, CC BY-SA, Open Government Licence – Toronto), with the attribution in LICENSES.md. CC BY-SA images remain under CC BY-SA.

The CC BY 4.0 grant for authored and generated images applies to the extent the authors hold rights in them.

To report a problem with any file, or to request its removal, open a discussion on this dataset page.

Citation

Please cite the Design Anything paper: https://arxiv.org/abs/2610.24771. Citation details are on the project page: https://designanything.gemus.ai