Despite perceptions of lagging behind in the artificial intelligence arena, Apple is quietly forging ahead with significant research. While Apple Intelligence represents a visible step, much of the company’s AI work happens behind the scenes, aiming to benefit the broader AI community, not just Apple users.
A recent project focuses on dramatically improving AI image editing through text prompts. Apple researchers unveiled Pico-Banana-400K, a meticulously curated dataset of 400,000 images designed to refine AI’s ability to understand and execute image modifications based on textual instructions.
The core problem Apple’s team addressed was the limitations of existing datasets. Many rely on AI-generated images, lacking the nuance of real-world photography, or suffer from a lack of diversity, hindering the development of truly versatile AI models.
Interestingly, Pico-Banana-400K is specifically designed to work in conjunction with Nano Banana, an image editing model developed by Google. This collaboration allows for the generation of 35 distinct types of edits, leveraging the power of Gemini-2.5-Pro to evaluate and refine the quality of those changes.
The dataset is structured with remarkable detail. It includes 258,000 examples comparing original images to their edited counterparts, 56,000 “preference pairs” highlighting successful versus failed edits, and 72,000 “multi-turn sequences” demonstrating a series of two to five consecutive edits.
Not all editing functions perform equally well. Broad stylistic changes and global edits prove relatively easy for the AI, achieving high success rates. However, tasks requiring precise geometric manipulation, understanding scene context, or accurate typography remain significantly more challenging.
The most successful function, artistic style transfer – transforming an image to resemble the style of Van Gogh or anime, for example – boasts a 93% success rate. Conversely, altering the font style or color of existing text only succeeds 58% of the time, revealing a critical area for improvement.
Other tested functions, like adding new text (67% success rate), zooming in (74% success rate), and applying vintage filters (91% success rate) fall between these extremes, providing a granular understanding of the AI’s capabilities and limitations.
In a departure from Apple’s typically closed ecosystem, Pico-Banana-400K is being released as an open resource for all researchers and AI developers. This commitment to open research is a notable move, particularly given Apple’s perceived position in the AI landscape.
While the future of AI-powered features like Siri remains uncertain, Apple’s dedication to advancing the underlying technology is undeniable. The company appears to be pursuing AI development on its own terms, focusing on foundational improvements that could benefit the entire field.
