Apple präsentiert KI-Datensatz für bessere Bildbearbeitung

Apple has published a new research paper demonstrating how AI image processing can become more precise and realistic in the future. The tech giant presents its own dataset of 400,000 images, designed to significantly improve the training basis for image processing AI. The project, named Pico-Banana-400K, is said to be one of the largest open datasets of its kind. With this dataset, Apple aims not only to improve its own systems but also to support the research community in developing more powerful and higher-quality image models.

Apple often seems to be lagging behind other major players in the field of artificial intelligence. While companies like OpenAI or Google attract attention with spectacular AI products, Apple tends to work more behind the scenes, focusing on the fundamentals.

One example of this is the research conducted by Apple Intelligence, which regularly presents new approaches and methods. In 2025, Apple published several studies, including one on AI's ability to detect programming errors. Now, another important step follows: a comprehensive image dataset intended to lay the foundation for improved, text-driven image processing systems.

What Apple wants to achieve with Pico-Banana 400K

In the study entitled „Pico-Banana-400K: A Large-Scale Dataset for Text-Guided Image Editing“, Apple describes how current AI systems deliver remarkable results in text-based image editing, but are limited by the quality and size of the training data used.

The researchers write that the progress of the entire research community continues to be limited by the lack of large, high-quality, and openly accessible datasets of real-world images. Existing datasets are often synthetically generated and inadequately reflect the diversity of real-world scenarios.

Pico-Banana-400K aims to close precisely this gap. The dataset contains approximately 400,000 real images that are freely available for non-commercial use. The goal is to train AI systems to better understand and implement image editing based on text instructions.

How the data set is structured

The dataset differs significantly from previous collections. It is organized according to a so-called 35-part editing taxonomy. This means that the images have been sorted according to typical types of editing – such as actions like moving, enlarging, inserting, changing colors, or adding artistic effects.

Each image is linked to a corresponding text instruction that describes what should happen during processing. This allows AI systems to learn not only to analyze images, but also how a desired change will actually affect them.

The edits range from simple actions – such as moving an object – to complex, multi-stage edits performed across multiple text instructions. The dataset also includes so-called preference pairs. These compare successful and unsuccessful edits so that AI models can recognize which results are considered erroneous.

How Apple ensures data quality

To ensure the data set met its own standards, Apple developed a multi-stage process. First, all input prompts were applied to the Nano-Banana system, which implemented the required image processing. Then, Gemini 2.5-Pro automatically checked the generated results.

Each image was individually evaluated and either accepted or rejected. This resulted in a dataset consisting of validated and consistent images. The entire process is designed to ensure that the training data is not only large but also reliable.

Where progress lies

The strength of Pico-Banana-400K lies not only in the sheer volume of data, but above all in its systematic structure. The combination of real-world images, clear text instructions, and the additional preference pairs allows AI models to recognize subtle differences and execute edits more precisely.

Apple highlights that the dataset is suitable for both simple and complex processing tasks, thus providing a solid foundation for the further development of AI systems in the field of image processing.

For research, this means that models will not only be more aesthetically pleasing in the future, but also more transparent and precise. Developers can freely use the dataset to train their own models or improve existing systems.

How Apple integrates research into its own products

Apple has consistently demonstrated in recent years its ability to gradually integrate research findings into its own products. For example, in June 2025, the company expanded its Image Playground, adding new image styles based on ChatGPT that users can apply in apps like Photos and Messages.

The work on Pico-Banana-400K could, in the long run, make such functions even more realistic, precise, and flexible. It exemplifies how Apple views artificial intelligence not as a short-term trend, but as a long-term development that relies on a solid foundation.

New AI browser ChatGPT Atlas: More than just a search

Apple prioritizes quality over speed in its AI research.

With the Pico-Banana 400K, Apple is making a clear statement in AI research. The company is focusing on what many others neglect: the quality of the training data. Instead of aiming for quick results, Apple is focusing on sustainable improvements that could have the greatest impact in the long run.

The new dataset provides an open, structured, and high-quality foundation upon which researchers and developers can build. With this, Apple is contributing to making AI image processing not only more powerful but also more reliable in the future – a step that demonstrates that Apple's innovative power is often quieter, but no less effective. (Image: Shutterstock / metamorworks)