Apple is making strides in artificial intelligence by releasing a new dataset aimed at enhancing AI-based image editing tools. The company introduced the Pico-Banana-400K, a collection of 400,000 “text-guided” images designed to improve the capabilities of image editors that operate based on textual prompts. This initiative reflects Apple’s ongoing commitment to AI development, even as it faces stiff competition from industry leaders like OpenAI, Google, and Meta.
The researchers at Apple have identified shortcomings in existing datasets, noting that many either rely on images generated by AI models or lack sufficient diversity. By addressing these issues, Apple believes its dataset will provide a more robust foundation for improving AI editing models. The Pico-Banana-400K dataset is unique in its approach, as it is intended to be compatible with Google’s image editing model, Nano Banana.
Insights from the Dataset
According to the research team, the Pico-Banana-400K dataset allows for the generation of 35 different types of edits. It also utilizes Gemini-2.5-Pro to evaluate the quality of these edits and determine which should be integrated into the overall dataset. The dataset consists of various components, including 258,000 single edits, 56,000 preference pairs that differentiate between successful and unsuccessful edits, and 72,000 multi-turn sequences that document the editing process across two to five edits.
The study highlights varying success rates for different editing functions. Global edits and stylization, for instance, achieved high success rates, while more intricate tasks such as precise geometry and typography presented challenges. The function with the highest success rate, strong artistic style transfer, which can transform an image’s style to mimic famous artists like Van Gogh, recorded a remarkable 93% success rate. In contrast, changing the font style or color of visible text succeeded only 58% of the time.
Open Access for Researchers
In a departure from its typical business practices, Apple has made the Pico-Banana-400K dataset open for use by all researchers and AI developers. This move signals a commitment to fostering collaboration in AI research, an area where Apple has traditionally been more reserved. The open nature of the dataset contrasts with many of Apple’s usual offerings, which are often exclusive to its own platforms.
While the timeline for advancements such as an AI-powered Siri remains uncertain, the release of Pico-Banana-400K showcases Apple’s earnest efforts in the AI domain. The initiative illustrates how Apple is working to carve out its niche in a competitive landscape, focusing on improving the tools that could shape the future of image editing. As the company continues to innovate, the implications for both consumers and professionals in creative fields may be significant.
