Unsplash Dataset

The Unsplash Dataset is an image dataset published by Unsplash for training and testing machine learning models. It is built from the Unsplash photo library and pairs images with the search activity and metadata generated as people and applications use the service, so models can learn from real visual search behaviour rather than labels alone. It is aimed at researchers and machine learning teams working on computer vision, image search, aesthetic quality and semantic understanding. Three kinds of signal stand out. Real visual searches let models go beyond object detection toward higher level concepts such as metaphor, sentiment and complex semantics. Keyword level popularity data supports work on image beauty and quality. Rich photographic metadata covers camera makes and models, lenses and focal lengths. Unsplash states that all data in the dataset relates to images and that it does not and will never include user level data. Full field definitions are in the developer documentation. Two editions are offered. The Lite edition is a 1.5 GB download containing 25,000 images, 25,000 keywords and 1,000,000 searches, and it is free for commercial and non commercial use. The Full edition is roughly 40 GB with more than 4,800,000 images, 7,000,000 keywords and 350,000,000 searches; access is by request, and Unsplash does not publish a price for it. The dataset is published on the Unsplash Developers site alongside the Unsplash API documentation, status page and changelog, which is where its field documentation and access requests live.