The decision stays yours

The photo is yours, and so is the edit. Models are good at the tedious part in the middle, like tracing around a person or working out what’s near and what’s far. They’re no substitute for deciding to take the picture, or deciding when an edit is done. We don’t collect your photos to train on, and we don’t generate new images from prompts.

Machine learning means a program learns from examples instead of following rules someone wrote by hand. AI is the broader term for software that does a job a person used to do. An LLM, or large language model, is the kind that works with words. Apple’s on-device language model and Claude are LLMs. Apple’s segmentation model and Depth Anything aren’t. They look at a photo and give back a shape, or a map of what’s near and far.

Tap something, draw across it, or drag a box around it. Apple’s segmentation model then follows the edges and traces around whatever looks like one object. It doesn’t need to know what the thing is called, so a dog and a lamppost work the same way. After that, Apple’s on-device language model suggests a name for the mask so it’s easy to find in the list. The name is sometimes wrong, but the outline is what matters for your edit. Hair, glass, and overlapping objects can still trip it up, and the edge tools help clean those up.

The model estimates distance the way you do when you look at a picture. There’s no second camera or measurement involved. It looks at the same clues you would, like what overlaps what, how big things are, and what goes soft in the distance. It ranks every pixel as nearer or farther rather than measuring meters, so a poster on a wall can fool it just like a painted hallway can fool you. Portrait and ProRAW photos sometimes already have depth from the camera, and we use that when it’s there. Otherwise the estimate runs on your device.

Two of them. The segmentation model behind Magic Select is part of Apple’s Vision framework, and the language model that names masks is part of Apple Intelligence. Both come with the system and are shared by every app that uses them, and your photo isn’t sent to Apple. Depth Anything V2 was made by the University of Hong Kong and TikTok, and Darkroom ships the version Apple publishes in its machine learning models. Claude is Anthropic’s, and it only works with text.

No. Magic Select, depth masks, and mask names all run on your device, and the photo stays there. We never receive it, store it, or train on it.

Yes. Depth Anything is built into Darkroom, so depth masks never need a connection. Magic Select needs one the first time you use it, while the system downloads its model. Mask names use a language model that’s already on your device. Ask and preset descriptions live on this website, so those do need a connection.

No. These models only give back a selection, or a name for it. What you export is your photo with your own adjustments.

Because the model is part of the system, not part of Darkroom. The first time any app needs it, the system downloads it once and shares it with every app after that. Depth Anything is the only model that ships inside Darkroom. Either way, your photo stays on your device.

You need it to export an edit that uses masks. You can make the mask and see the result without a membership, and Foreground and Background on Portrait photos are free. There’s more in Masks & Local Adjustments.

Claude never sees a photo. For a shared preset, it reads the settings, like the sliders and curves, and writes a single sentence. That way nobody can fill the library with junk descriptions. For Ask, it reads our help pages and answers from them, and if the pages don’t cover something, it’s supposed to say so. The other language model is Apple’s. It runs on your device and only looks at your photo to name a Magic Select mask.