Apple's New AI Dataset Aims to Improve Photo Editing Models

Apple researchers have introduced Pico-Banana-400K, a dataset of 400,000 images designed to enhance AI photo editing based on text prompts. The dataset aims to address the current gap in AI image editing training, which has been limited by inadequate training data. Apple's new dataset features images organized into 35 different edit types across eight categories, including basic adjustments and complex transformations. Each image was evaluated using Apple's AI-powered quality control system and Google's Gemini-2.5-Pro. The dataset includes three specialized subsets for basic training, preference pairs, and multi-turn sequences. The subsets contain 258,000 single-edit examples, 56,000 preference pairs, and 72,000 multi-turn sequences. Apple built the dataset using Google's Gemini-2.5-Flash-Image editing model, which was released a few months ago. However, Apple's research revealed the limitations of this model, particularly in precise tasks like relocating objects or editing text. The success rates for these tasks were below 60%, while global style changes succeeded 93% of the time. The release of Pico-Banana-400K is expected to improve the performance of AI systems in editing photos based on text prompts.