Apple’s latest research paper argues that today’s AI image editors rely on inadequate training datasets—so Apple Intelligence researchers have created a new, improved one.
Despite the common perception that Apple lags behind in artificial intelligence, the company continues to publish substantial research in the field. In 2025 alone, it has released influential studies exploring how AI can detect code bugs and exposing its limitations in reasoning.
Now, Apple has unveiled another major work: “Pico-Banana-400K: A Large-Scale Dataset for Text-Guided Image Editing.” The paper focuses on improving how AI systems learn to edit images based on text prompts.
While the researchers praise existing models like GPT‑4o and Nano‑Banana as “remarkable” in text-guided image editing, they identify a core limitation in current approaches. According to the paper, “the research community’s progress remains constrained by the absence of large-scale, high-quality, and openly accessible datasets built from real images.”
To address that, Apple’s team developed Pico‑Banana‑400K, a collection of 400,000 high-quality images created specifically for instruction-based image editing. The dataset is freely available for non-commercial use and organized by a 35-category editing taxonomy, representing the types of changes users typically request—such as moving objects, applying artistic effects, or adjusting framing and zoom.
How the dataset was built
Each image in Pico‑Banana‑400K was uploaded to Nano‑Banana along with a textual editing prompt. Then Apple researchers used Gemini‑2.5‑Pro to analyze the results, filtering out unsatisfactory outputs until only approved images remained.
The final dataset includes single-turn edits (one prompt per edit), multi-turn sequences (progressive edits across prompts), and preference pairs—side-by-side comparisons of successful versus failed image outcomes. These pairs help AI models learn not only what “good” results look like, but also how to avoid producing bad ones.
Why it matters
By open-sourcing Pico‑Banana‑400K, Apple aims to provide a stronger foundation for building the next generation of AI image editors. The company says this dataset “establishes a robust foundation for training” systems capable of understanding and executing complex, text-driven image edits with precision and creativity.
Separately, Apple also expanded its own Image Playground tool in June 2025, adding a wider range of ChatGPT-powered image styles—underscoring how its research continues to translate into practical advances for users.
