Apple has launched Pico-Banana-400K, a curated 400,000-image analysis dataset which, apparently, was constructed utilizing Google’s Gemini-2.5 fashions. Listed here are the main points.

Apple’s analysis crew has printed an fascinating examine known as “Pico-Banana-400K: A Massive-Scale Dataset for Textual content-Guided Picture Modifying”.

Along with the examine, additionally they launched the complete 400,000-image dataset it produced, which has a non-commercial analysis license. Which means that anybody can use it and discover it, offered it’s for educational work or AI analysis functions. In different phrases, it may’t be used commercially.

Proper, however what’s it?

Just a few months in the past, Google launched the Gemini-2.5-Flash-Picture mannequin, also called Nanon-Banana, which is arguably the state-of-the-art in relation to picture enhancing fashions.

Different fashions have additionally proven vital enhancements, however, as Apple’s researchers put it:

“Regardless of these advances, open analysis stays restricted by the dearth of large-scale, high-quality, and absolutely shareable enhancing datasets. Present datasets usually depend on artificial generations from proprietary fashions or restricted human-curated subsets. Moreover, these datasets steadily exhibit area shifts, unbalanced edit sort distributions, and inconsistent high quality management, hindering the event of strong enhancing fashions.”

So, Apple got down to do one thing about it.

Constructing Pico-Banana-400K

The very first thing Apple did was pull an unspecified variety of actual images from the OpenImages dataset, “chosen to make sure protection of people, objects, and textual scenes.”

Then, it got here up with a listing of 35 various kinds of modifications a person might ask the mannequin to make, grouped into eight classes. For example:

• Pixel & Photometric: Add movie grain or classic filter
• Human-Centric: Funko-Pop–fashion toy determine of the particular person
• Scene Composition & Multi-Topic: Change climate circumstances (sunny/wet/snowy)
• Object-Stage Semantic: Relocate an object (change its place/spatial relation)
• Scale: Zoom in

Subsequent, the researchers would add a picture to Nano-Banana, alongside one in every of these prompts. As soon as Nano-Banana was executed producing the edited picture, the researchers would then have Gemini-2.5-Professional analyze the consequence, both approving it or rejecting it, based mostly on instruction compliance and visible high quality.

The consequence grew to become Pico-Banana-400K, which incorporates photos produced via single-turn edits (a single immediate), multi-turn edit sequences (a number of iterative prompts), and desire pairs evaluating profitable and failed outcomes (so fashions may be taught what undesirable outcomes appear to be).

Whereas acknowledging Nano-Banana’s limitations in fine-grained spatial enhancing, structure extrapolation, and typography, the researchers say that they hope Pico-Banana-400K will function “a strong basis for coaching and benchmarking the following era of text-guided picture enhancing fashions.”

Yow will discover the examine on arXivand the dataset is freely obtainable on GitHub.

Accent offers on Amazon