Abstract
We introduce ArtBench-10, the first class-balanced, high-quality, cleanlyannotated, and standardized dataset for benchmarking artwork generation. Itcomprises 60,000 images of artwork from 10 distinctive artistic styles, with5,000 training images and 1,000 testing images per style. ArtBench-10 hasseveral advantages over previous artwork datasets. Firstly, it isclass-balanced while most previous artwork datasets suffer from the long tailclass distributions. Secondly, the images are of high quality with cleanannotations. Thirdly, ArtBench-10 is created with standardized data collection,annotation, filtering, and preprocessing procedures. We provide three versionsof the dataset with different resolutions ($32\times32$, $256\times256$, andoriginal image size), formatted in a way that is easy to be incorporated bypopular machine learning frameworks. We also conduct extensive benchmarkingexperiments using representative image synthesis models with ArtBench-10 andpresent in-depth analysis. The dataset is available athttps://github.com/liaopeiyuan/artbench under a Fair Use license.