ArtSem

The dataset introduced in "Controllable Multi-domain Semantic Artwork Synthesis" (Computational Visual Media, Volume 10, pages 355–373, 2024).

ArtSem contains 40,000 artwork images across four domains — ink-wash, Monet oil, Van Gogh oil, and watercolor — each paired with a semantic label map.

• Paper: https://doi.org/10.1007/s41095-023-0356-2
• Project page: https://sky24h.github.io/websites/cvmj2024_controllable-artwork-synthesis/
• Code: https://github.com/sky24h/Controllable_Multi-domain_Semantic_Artwork_Synthesis
• Demo: https://huggingface.co/spaces/sky24h/Controllable_Multi-domain_Semantic_Artwork_Synthesis

Usage

from datasets import load_dataset

ds = load_dataset("sky24h/ArtSem", split="train")          # all four domains
ds = load_dataset("sky24h/ArtSem", "ink", split="train")   # one domain

Each sample has three fields:

Field	Description
image.png	The artwork, 512x512 RGB.
mask.png	The semantic label map, 512x512 single-channel index map with values 0-15.
json	{"domain": ..., "id": ...}, where id is the source photograph's Flickr id.

The label map is an index map, not a color image; opening it as RGB or applying a palette will change the values. To stream instead of downloading all 19 GB, pass streaming=True.

Contents

40,000 train pairs, 10,000 per domain, packed as WebDataset shards of about 1 GB under data/ — 20 train shards and one test shard per domain:

data/train-<domain>-<nnnn>.tar
data/test-<domain>-0000.tar

The 16 classes of the label maps are:

0 building	1 clouds	2 dirt	3 grass	4 ground	5 hill	6 mountain	7 plant
8 river	9 road	10 rock	11 sea	12 sky	13 snow	14 tree	15 other

Per-class statistics are given in Fig. 8 of the paper's supplementary material. The display colors used by the interactive demo are listed as CLASSES in inference.py in the code repository; they are a visualization convention only and are not part of the data.

Splits

The paper does not define a held-out test split. Following the usual protocol for generative models, FID is computed between generated artworks and the real artworks of the dataset, so all 40,000 pairs are used for training.

The test split is a convenience subset of 100 pairs per domain, taken verbatim from train, and is intended as a small fixed set of layouts to run inference on. It is not held out from training. If you need a held-out evaluation set, split train yourself.

Cross-domain structure

The four domains were selected independently — for each domain, the 10,000 highest-scoring generated artworks were kept — so their source photographs overlap only partially. Across the 40,000 pairs there are 22,460 distinct source photographs:

appears in	1 domain	2 domains	3 domains	4 domains
photographs	9,780	8,429	3,642	609

Where the same id appears in more than one domain, the label map is identical and only the artwork differs, giving cross-domain pairs that share a layout.

Each shard holds a single domain, so shuffle across shards if you need domain-mixed batches.

How the data was produced

Label maps were derived from landscape photographs collected from Flickr, and the artworks were then generated from them by a weakly supervised generation stage. See Section 3 of the paper for the full construction pipeline.

The artwork images are therefore synthesized, not photographs of real paintings, and no source photograph is redistributed here.

License and intended use

Released for non-commercial academic research use only, under CC BY-NC-SA 4.0, matching the accompanying code.

Contact

If you find any issue with the dataset, please contact the author at hytian2 [at] gmail [dot] com.

Citation

@Article{Huang2024CMSAS,
  author={Yuantian Huang and Satoshi Iizuka and Edgar Simo-Serra and Kazuhiro Fukui},
  title={Controllable Multi-domain Semantic Artwork Synthesis},
  journal={Computational Visual Media},
  volume={10},
  number={2},
  pages={355--373},
  year={2024},
  publisher={Springer},
  url={https://doi.org/10.1007/s41095-023-0356-2}
}