Apple could be useless final within the AI race—a minimum of when you think about competitors from corporations like OpenAI, Google, and Meta—however that does not imply the corporate is not engaged on the tech. The truth is, it appears a lot of the work Apple does on AI is behind the scenes: Whereas Apple Intelligence is, properly, there, the corporate’s researchers are engaged on different methods to enhance AI fashions for everybody, not simply Apple customers. The most recent venture? Bettering AI picture editors based mostly on textual content prompts.
In a paper printed final week, researchers launched Pico-Banana-400K, a dataset of 400,000 “text-guided” photos chosen to enhance AI-based picture enhancing. Apple believes its picture dataset improves upon current units by together with increased high quality photos with extra range: The researchers discovered that current datasets both use photos produced by AI fashions, or aren’t diversified sufficient, which may hinder efforts to enhance the fashions.
Funnily sufficient, Pico-Banana-400K is designed to work with Nano Banana, Google’s picture enhancing mannequin. Researchers say utilizing Nano Banana, their dataset can generate 35 several types of edits, in addition to faucet into Gemini-2.5-Professional to asses high quality the edits, and whether or not these edits ought to stay as a part of the general dataset.
As a part of these 400,000 photos, there are 258,000 samples of single edits (the place Apple compares the unique photos to at least one with edits); 56,000 “desire pairs,” which distinguishes between failed and profitable edit generations; and 72,000 “multi-turn sequences,” which walks by way of two to 5 edits.
What do you assume thus far?
Researchers be aware that totally different capabilities had totally different success charges on this dataset. World edits and stylization are “simple,” attaining the best success charges; object semantics and scene context are “reasonable;” whereas exact geometry, format, and typography are “arduous.” The very best performing operate, “robust inventive model switch,” which may embrace altering a picture’s model to “Van Gogh” or anime, has a 93% success charge. The bottom performing operate, “change font model or shade of seen textual content if there’s textual content,” solely succeeded 58% of the time. Different examined capabilities embrace “add new textual content” (67% success charge), “zoom in” (74% success charge), and “add movie grain or classic filter” (91% success charge).
Not like lots of Apple’s merchandise, that are usually closed to the corporate’s personal platforms, Pico-Banana-400K is open for all researchers and AI builders to make use of. It is cool to see Apple researchers contributing to open analysis like this, particularly in an space Apple is mostly behind in. Will we truly get an AI-powered Siri anytime quickly? Unclear. However it’s clear Apple is actively engaged on AI, maybe simply in its personal means.
