ArtBench-10 is a class-balanced, high-quality, cleanly annotated, and standardized dataset for benchmarking artwork generation. The dataset includes 60,000 artwork images from 10 unique art styles, with 5,000 training images and 1,000 test images for each style.

Compared with previous artwork datasets, ArtBench-10 has several advantages:

• First, it is class-balanced, while most previous artwork datasets have the problem of long-tailed class distribution.
• Second, the images are of high quality and the annotations are clean.
• Third, ArtBench-10 was created through standardized data collection, annotation, filtering, and preprocessing procedures.

The authors provide three versions of the dataset with different resolutions (32×32, 256×256, and original image size), formatted in a way that can be easily adopted by popular machine learning frameworks.