How Apple Plans to Improve AI Image Editors

Apple may be dead last in the AI ​​race – at least when you consider competition from companies like OpenAI, Google and Meta – but that doesn’t mean the company isn’t working on the technology. In fact, it seems like most of the work Apple does on AI happens behind the scenes: While Apple Intelligence is there, the company’s researchers are working on other ways to improve AI models for everyone, not just Apple users. The latest project? Improved AI image editors based on text prompts.

In a paper published last week, researchers presented Pico-Banana-400K, a dataset of 400,000 “text-guided” images selected to improve AI-based image editing. Apple believes its image dataset improves on existing ones by including higher quality and more diverse images: Researchers have found that existing datasets use images produced by AI models or are not diverse enough, which can hamper efforts to improve the models.

Oddly enough, Pico-Banana-400K is designed to work with Nano Banana, Google’s image editing model. The researchers say that using Nano Banana, their dataset can generate 35 different types of edits, as well as tap into Gemini-2.5-Pro ​​to assess the quality of edits and whether those edits should remain in the overall dataset.

As part of those 400,000 images, there are 258,000 unique edit samples (where Apple compares the original images to an image with edits); 56,000 “preference pairs,” which distinguish between failed and successful edit generations; and 72,000 “multi-turn sequences,” which run through two to five edits.

The researchers note that different functions had different success rates in this dataset. Overall edits and stylization are “easy,” achieving the highest success rates; the semantics of the objects and the context of the scene are “moderate”; while precise geometry, layout, and typography are “difficult.” The best-performing feature, “strong art style transfer,” which could include changing the style of an image to “Van Gogh” or anime, has a 93% success rate. The worst-performing feature, “change the font style or color of visible text if there is text,” only succeeded 58% of the time. Other functions tested include “add new text” (67% success rate), “zoom” (74% success rate), and “add film grain or vintage filter” (91% success rate).

Unlike many Apple products, which are generally closed to the company’s own platforms, Pico-Banana-400K is open to all AI researchers and developers. It’s cool to see Apple researchers contributing to open research like this, especially in an area where Apple is usually behind. Will we soon have an AI-powered Siri? Unclear. But it’s clear that Apple is actively working on AI, perhaps just in its own way.