A new Apple research paper argues that AI imaging editors are currently trained on inadequate image sets — so Apple Intelligence researchers have released an improved one.
Despite the continual presumption that Apple is behind the industry in AI, it keeps publishing comprehensive research papers on the subject. In 2025 alone, it's most significant studies have covered how AI cannot reason, but can uncover bugs in code.
Now the researchers have published "Pico-Banana-400K: A Large-Scale Dataset for Text-Guided Image Editing." It's explicitly concerned with how to better train AI systems to edit images following text prompts.
Despite describing current systems like GPT-4o and Nano-Banana being "remarkable [at] text-guided image editing," the paper claims that there is a key limitation with how they all work.
"[The] research community's progress remains constrained by the absence of large-scale, high-quality, and openly accessible datasets built from real images," says the full paper.
So Apple's researchers have launched "Pico-Banana-400K, a comprehensive 400K-image dataset for instruction-based image editing." As well as being a large set, what "distinguishes Pico-Banana-400K from previous synthetic datasets is our systematic approach to quality and diversity."
Example from the research paper showing the training process described — image credit: Apple
The approximately 400,000 images in the set have all been made freely available for non-commercial use. They are "organized by a 35-type editing taxonomy," meaning types of image edits a user could typically want.
What the researchers did
Those include edits such as moving an object in the image, adding artistic effects, and zooming on. Apple's researchers uploaded each image in the set to Nano-Banana, together with such a prompt.
Using Gemini-2.5-Pro, the researchers had the resulting images analyzed and then either rejected or accepted them.
"The result became Pico-Banana-400K, which includes images produced through single-turn edits (a single prompt), multi-turn edit sequences (multiple iterative prompts)," say the researchers, "and preference pairs comparing successful and failed results (so models can also learn what undesirable outcomes look like)."
Having now produced this large image dataset, Apple's researchers say Pico-Banana-400K "establishes a robust foundation for training" AI image editors.
Separately, Apple most recently improved its own Image Playground in June 2025. It added more ChatGPT-powered image styles.
