Apple releases 400K-image dataset to train AI editing models

Apple has released Pico-Banana-400K, a 400,000-image research dataset designed to train AI image editing models, built using Google’s Gemini-2.5 technology. The dataset addresses a critical gap in open AI research by providing high-quality, shareable training data that researchers say has been limited by synthetic generations from proprietary models and inconsistent quality control.

Why this matters: Existing image editing datasets often suffer from domain shifts, unbalanced edit distributions, and quality issues that hinder the development of robust AI models, leaving researchers without adequate training resources.

How Apple built the dataset: The research team systematically created the dataset using a multi-step validation process to ensure quality and diversity.

• Apple pulled real photographs from the OpenImages dataset, selecting images that covered humans, objects, and textual scenes.
• Researchers developed 35 different types of editing prompts grouped into eight categories, including pixel & photometric changes (add film grain), human-centric edits (Funko-Pop–style transformations), and scene composition modifications (weather changes).
• Each image was processed through Google’s Nano-Banana (Gemini-2.5-Flash-Image) model with specific prompts, then analyzed by Gemini-2.5-Pro for approval based on instruction compliance and visual quality.

What’s included: The dataset encompasses multiple types of AI training scenarios to help models learn both successful and unsuccessful editing outcomes.

• Single-turn edits using individual prompts for immediate transformations.
• Multi-turn edit sequences involving multiple iterative prompts for complex modifications.
• Preference pairs comparing successful and failed results so models can learn to identify undesirable outcomes.

The bigger picture: Apple’s researchers acknowledge that current image editing models, including Nano-Banana, still struggle with fine-grained spatial editing, layout extrapolation, and typography challenges.

• The team hopes Pico-Banana-400K will serve as “a robust foundation for training and benchmarking the next generation of text-guided image editing models.”
• The dataset is available under a non-commercial research license, meaning it can be used for academic work and AI research but not commercial applications.

Where to find it: The complete study is available on arXiv, with the full dataset accessible through GitHub for qualifying researchers.