Apple’s Pico-Banana Dataset Aims to Shake Up Photo Editing

Apple has rolled out an open research dataset named Pico-Banana-400K, which is geared toward enhancing the capability of AI models to edit photos through text-based instructions. This development may represent one of Apple’s most significant steps forward in AI research in recent years.

A Resource for Better Image Interpretation

This dataset is packed with 400,000 real images, each paired with AI-generated edits, crafted using Apple’s internal Nano-Banana model. In each case, a picture is altered based on prompts such as “add a rainbow” or “blur the background.” Through this, Apple aims to train AI models in interpreting natural language instructions and conducting realistic image modifications.

By utilizing real photos from the OpenImages collection instead of synthetic ones, Pico-Banana-400K aims to close the gap between the data used in AI training and actual image quality, a critical step in refining AI capabilities.

Constructed with Careful Attention to Detail

Apple organized this dataset around 35 distinct edit types, touching on changes in color, lighting, object removal, background modifications, and more. To ensure the accuracy of each edit, Google’s Gemini 2.5 Pro was used to automatically assess how well the edits matched the text prompts and their technical quality. Edits that didn’t meet standards were reworked until satisfactory.

The dataset is broken down into three investigative subsets:

• Single edits amount to 258,000 examples—simple, one-step modifications.
• Multi-turn edits total 72,000—showcasing sequential transformations.
• Preference pairs feature 56,000 comparative examples of good vs. poor edits to refine model decisions.

Strengthening AI Model Limitations

Testing shows that current AI models handle general edits like tone shifts well, but face challenges with specific tasks like moving objects or editing text within images. By providing a mix of success and failure examples, Apple is looking to guide researchers in training models that better understand context and can rectify errors.

Available for Research, Not Commercial Ventures

The Pico-Banana-400K can be accessed through GitHub, offered under a non-commercial license. This collection opens new possibilities for developers aspiring to create and assess the next wave of text-driven image editors. Apple’s move is a deliberate attempt to engage more significantly with the AI research community while also likely enhancing features such as Image Playground and potential Apple Intelligence applications.

With Pico-Banana-400K, Apple indicates a gradual yet notable transition from a traditionally closed AI approach to one of more open, collaborative efforts in evolving visual AI technologies.

Is this a development you’re interested in? Share your thoughts below.