![\"Apple\u2019s](\"https:\/\/dataconomy.com\/wp-content\/uploads\/2025\/11\/apple-pico-banana-400k.jpg\" "\\"Apple\u2019s")
<\/p>\n<\/p>\n
Apple has released Pico-Banana-400K<\/strong>, a massive, high-quality dataset of nearly 400,000 image editing examples. The new dataset, detailed in an academic paper posted on October 23, 2025, was built by Apple researchers including Yusu Qian, Jialing Tong, and Zhe Gan. This matters because the AI community has been held back by a lack of large-scale, open, and realistic datasets. Most previous datasets were either synthetic, low-quality, or built with proprietary models. Apple\u2019s new resource, which is built from real photographs, is designed to be a robust foundation for training the next generation of text-guided image editing models, from simple touch-ups to complex, multi-step creative projects.<\/p>\n Instead of the old, expensive method of paying humans to manually edit hundreds of thousands of images, Apple\u2019s team created a sophisticated, automated pipeline using other powerful AI models. . First, they sourced real photographs from the OpenImages<\/strong> collection. Then, they used Google\u2019s Nano-Banana model to generate a diverse range of edits based on a comprehensive taxonomy of 35 different edit types, from \u201cchange color\u201d to \u201capply seasonal transformation.\u201d<\/p>\n But here\u2019s the clever part: to ensure quality, they used another AI, Gemini-2.5-Pro<\/strong>, as an automated \u201cjudge.\u201d This AI judge scored every single edit on four criteria: Instruction Compliance (40%)<\/strong>, Seamlessness (25%)<\/strong>, Preservation Balance (20%)<\/strong>, and Technical Quality (15%)<\/strong>. Edits that scored above a 0.7 threshold were labeled \u201csuccessful.\u201d Edits that failed were kept as \u201cnegative examples.\u201d This process creates a high-quality dataset without a single human annotator, at a total cost of about $100,000.<\/p>\n The real power of Pico-Banana-400K isn\u2019t just its size; it\u2019s the specialized subsets designed to solve complex research problems. The full dataset includes:<\/p>\n By analyzing the \u201csuccess rates\u201d of its own pipeline, the Apple team also created a clear map of what AI image editors are good at and where they still fail. Global edits like \u201cadd a vintage filter\u201d (90% success) are easy. Object-level edits like \u201cremove this car\u201d (83% success) are pretty good. But edits requiring precise spatial control or symbolic understanding remain \u201cbrittle\u201d and are now open problems for researchers to solve.<\/p>\n The hardest tasks? Relocating an object<\/strong> (59% success), changing a font<\/strong> (57% success), and generating caricatures<\/strong> (58% success). By open-sourcing this dataset, Apple is essentially giving the entire AI community a high-quality \u201cgym\u201d to train their models and a clear list of challenges to tackle next.<\/p>\nHow Pico-Banana-400K was built<\/h2>\n
More than just single edits<\/h2>\n
\n
What this means for future AI editors<\/h2>\n