Recently, Apple has made public a substantial dataset called Pico-Banana-400K. This dataset encompasses 400,000 meticulously curated images, with the overarching goal of propelling the progress of text-guided image editing technology.

Leveraging the capabilities of models from the Google Gemini-2.5 series, the dataset is segmented into three distinct components: a single-round supervised fine-tuning (SFT) subset, a multi-round editing set, and a preference set.

The single-round SFT subset boasts 258,000 instances of successful single-round image editing, spanning across 35 different editing categories. The multi-round editing set, on the other hand, features 72,000 multi-round editing sequences. These sequences are tailored for in-depth exploration of sequential editing, logical reasoning, and strategic planning in the context of continuous image modifications.

The preference set rounds out the dataset with 56,000 examples. Each example is a triplet consisting of an original image, editing instructions, a successful edit, and an unsuccessful edit. This set is specifically crafted to facilitate alignment research.

In addition, Apple has developed an automated editing and evaluation pipeline. In this pipeline, the Nano-Banana model takes charge of the editing tasks, while the Gemini-2.5-Pro model is responsible for evaluating the outcomes. For cases where the editing fails, the system automatically retries the process until a successful result is achieved, all without the need for human intervention.

The research paper associated with this dataset, titled "Pico-Banana-400K: A Large-Scale Dataset for Text-Guided Image Editing," has been published concurrently.