GPIC, short for Giant Permissive Image Corpus, is a vast visual dataset designed to accelerate scalable generative model research and development.

An arXiv paper (2605.30341v1) provides the scholarly foundation for GPIC, detailing its methodology.

Researchers have established a standardized benchmarking protocol for GPIC, including a reference baseline for pixel-space flow matching to enable immediate use and apples-to-apples comparisons.

The dataset encompasses roughly 28 trillion pixels across 100 million training examples, 200,000 validation samples, and 1 million test examples, and includes state-of-the-art vision-language captions.

All GPIC images are permissively licensed, removing licensing barriers for both academic research and commercial deployment.

The initiative is aimed at unlocking scale, democratizing access to large-scale visual data, and speeding the development of next-generation generative AI models.