Apple Just Released 400,000 Photos to Help AI Edit Images Better

| October 28, 2025

If you’ve ever asked an AI to “make this photo look like a cartoon” or “change the sky to sunset” and gotten weird results—you’re not alone. Now, Apple is trying to fix that.

This week, Apple researchers released a huge new collection of photos called Pico-Banana-400K—a free dataset with 400,000 carefully chosen images designed to help train AI systems to edit photos more accurately using text commands (like “turn this person into a LEGO figure” or “brighten the colors”).

Why This Matters

Right now, many AI photo editors struggle with simple tasks. They might change the whole mood of a photo beautifully—but fail if you ask them to move a tree or fix a typo on a sign. Why? Because they haven’t been trained on enough real, high-quality examples of good edits.

Apple says this has been a big problem in AI research—so they built a better training set.

What’s Special About This Dataset?

Unlike older collections, Pico-Banana-400K is super organized and carefully checked:

• It covers 35 types of edits across 8 categories—from basic color fixes to fun transformations like “make this look like a Pixar movie.”

• Every image was created and reviewed by powerful AI models (Google’s Gemini-2.5 systems) to make sure the edits actually match the instructions.

• Nothing was added by chance—each photo passed a strict quality test before making the final cut.

The dataset also includes three helpful parts:

1. 258,000 single edits (great for basic AI training)

2. 56,000 pairs showing a good edit vs. a bad one (so AI can learn what works)

3. 72,000 multi-step edits (like “first make it sunny, then add sunglasses”)

What Apple Learned

While testing, Apple found that AI is great at big, general changes—like switching a photo to black-and-white or giving it a watercolor look (success rate: 93%).
But it struggles with small, precise tasks, like moving an object or editing text in an image (success rate: under 60%).

That explains why your AI might turn your dog into a dragon—but can’t move it to the left side of the picture!

Free for Researchers

The best part? Apple is sharing the entire dataset for free on GitHub—for non-commercial research only. That means universities, startups, and developers can use it to build smarter, more reliable photo-editing AI.

As Apple’s team put it: this dataset is a “strong foundation” for the next generation of AI that actually understands what you mean when you say, “Make it pop!”