Apple has quietly released a new research dataset called Pico-Banana-400K: A Large-Scale Dataset for Text-Guided Image Editing. It is said to be one of its biggest open contributions to the AI research community in years. The research paper was published on arXiv and is designed to advance research in instruction-based image editing.
It includes 400,000 real-world images sourced from the OpenImages collection, paired with artificially edited counterparts. These edits were generated using the Nano-Banana model to produce diverse, high-quality edit examples. Apple’s researchers say this approach helps cover a wide range of visual transformations while maintaining strong consistency between text instructions and image outputs.
In order to ensure depth and quality, the company built the dataset around a detailed taxonomy of 35 different edit types, spanning color changes, object addition or removal, background alteration, and other complex adjustments. Each image-edit pair underwent quality scoring by multimodal large language models (MLLMs) and further human curation to minimise noise and maintain reliability.
The dataset comes with three key subsets for different types of research. One includes 72,000 multi-turn examples that focus on sequential editing, reasoning, and planning across multiple steps. Another offers 56,000 preference-based examples meant for alignment and reward model training. There’s also a set of paired long and short instructions that can help researchers study how AI models rewrite or summarise text commands effectively.
Furthermore, Apple’s Pico-Banana-400K also attempts to fix some long-standing issues in text-to-image research, like over-reliance on synthetic data, domain shifts between datasets, and inconsistent quality control. The inclusion of both successful and failed edit examples gives models a chance to learn from mistakes.
Available for non-commercial academic and AI research use, the dataset offers a strong foundation for training and benchmarking the next generation of text-guided image editing systems.
