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

Apple researchers person released Pico-Banana-400K, a broad dataset of 400,000 curated images that's been specifically designed to amended however AI systems edit photos based connected substance prompts.


The monolithic dataset aims to code what Apple describes arsenic a spread successful existent AI representation editing training. While systems similar GPT-4o tin marque awesome edits, the researchers accidental advancement has been constricted by inadequate grooming information built from existent photographs. Apple's caller dataset aims to amended the situation.

Pico-Banana-400K features images organized into 35 antithetic edit types crossed 8 categories, from basal adjustments similar colour changes to analyzable transformations specified arsenic converting radical into Pixar-style characters oregon LEGO figures. Each representation went done Apple's AI-powered prime power system, with Google's Gemini-2.5-Pro being utilized to measure the results based connected acquisition compliance and method quality.

The dataset besides includes 3 specialized subsets: 258,000 single-edit examples for basal training, 56,000 penchant pairs comparing palmy and failed edits, and 72,000 multi-turn sequences showing however images germinate done aggregate consecutive edits.

Apple built the dataset utilizing Google's Gemini-2.5-Flash-Image (aka Nano-Banana) editing model, which was released conscionable a fewer months ago. However, Apple's probe revealed its limitations. While planetary benignant changes succeeded 93% of the time, precise tasks similar relocating objects oregon editing substance earnestly struggled, with occurrence rates beneath 60%.


Despite the limitations, researchers accidental their purpose with Pico-Banana-400K is to found "a robust instauration for grooming and benchmarking the adjacent procreation of text-guided representation editing models." The complete dataset is freely available for non-commercial probe usage connected GitHub, truthful developers tin usage it to bid much susceptible representation editing AI.
This article, "Apple's New AI Dataset Aims to Improve Photo Editing Models" archetypal appeared connected MacRumors.com

Discuss this article successful our forums