Apple provides training data for AI image editing with Google's Nano Banana
News from Apple's AI research department: With Pico-Banana-400K and Google's Gemini-2.5-Pro along with Nano Banana, images can be edited better.
Usage workflow of Pico-Banana-400K.
(Image: Apple)
Apple's research department has presented a new package of training data that is intended to help AI image generators perform better edits. Interestingly, “Pico-Banana-400K” builds on Google's Gemini-2.5-Pro along with the popular image generator Nano Banana from the internet giant and has been verified with it. The dataset contains, as the name suggests, nearly 400,000 text-image-edit triplets. The goal is to facilitate research into text-controlled image manipulation. This means: if you want to edit an existing image using a prompt, the respective model should understand the user better.
Gift to Research
According to the paper by Apple researchers, Pico-Banana-400K allows for a total of 35 different edit operations across eight semantic categories. It is said to cover almost everything known from Photoshop and beyond, such as color changes down to the low-level area, style changes, or the reconstruction of entire objects and scenes.
Videos by heise
The motivation for the dataset, according to the researchers, is the lack of “large, high-quality, and freely accessible datasets of real images” suitable for such models. “Recent advances in multimodal models have shown remarkable text-guided image editing capabilities, with systems like GPT-4o and Nano-Banana setting new benchmarks,” the researchers write. With Pico-Banana-400K, even more should now be possible.
For Non-Commercial Use Only
Pico-Banana-400K is a “robust foundation for training and benchmarking” new models, the researchers continue. The package is hosted on Apple's servers and can be downloaded component by component via manifest files. Something unfortunate: Apple does not foresee commercial use. Instead, the company relies on a CC-BY-NC-ND 4.0 license. This means: non-commercial applications and research are allowed free of charge, but not used for paid projects.
The company also prohibits the redistribution of derivatives of the dataset. The images themselves are all CC-BY 2.0 materials, meaning commercially usable images but requiring attribution. Users of the dataset must commit to complying with both licenses, although attribution is known to be very difficult in the AI field. How this should be handled with newly generated images is unclear.
Empfohlener redaktioneller Inhalt
Mit Ihrer Zustimmung wird hier ein externer Preisvergleich (heise Preisvergleich) geladen.
Ich bin damit einverstanden, dass mir externe Inhalte angezeigt werden. Damit können personenbezogene Daten an Drittplattformen (heise Preisvergleich) übermittelt werden. Mehr dazu in unserer Datenschutzerklärung.
(bsc)
