Pico-Banana-400k has been released by researchers from Apple. It is an expansive dataset of 400,000 images that have been compiled and designed to improve how AI systems perform the editing of pictures that are based from text prompts.
The big dataset is made to deal with what Apple currently calls a gap in the editing and training of images in AI systems. For instance, GPT-4o can create great edits that can impress people. The researchers state that the inadequacy of training data based on real photos hinders the progress.
The Pico-Banana-400k has images that are compiled into 35 edit types spanning eight categories, including complex changes and simple color adjustments. Examples are where people are turned into LEGO figures or are painted in Pixar-style. Even with the limitations, the researchers state that Pico-Banana-400k will be a solid foundation for benchmarking and training. Developers can access the dataset on GitHub.
