Opening a photo
When you drop or choose a picture, the browser reads the file from your disk into the page's memory. JPG, PNG, WebP, AVIF, GIF and BMP are decoded by the browser itself; HEIC and TIFF decoders load only when a file of that type is opened. A working copy of up to 1,600 pixels on the long side drives the preview, and the original stays in memory for the export.
Finding the subject
Detection runs with ONNX Runtime Web, compiled to WebAssembly and served from this site, using the graphics processor through WebGPU where the browser allows it and the processor otherwise. Auto first runs a 0.2 MB face finder. If it sees a face it uses the person model (26 MB), otherwise the object model (84 MB). The model files come from our model file server the first time and are kept in this browser's cache, so later cut-outs, and cut-outs without a connection, start immediately. A model contains no pictures; it is a set of numbers the detector uses.
Several jobs need no model at all: removing a white or black background, logos, signatures, the hand eraser and adding a background to a transparent picture. They run the moment the picture opens.
A document, not a pile of pixels
Everything you change after the cut-out, from the background and canvas to the shadow and edge settings, is stored as a small description of the result rather than as edited pixels. That is why switching between twenty backgrounds costs nothing, why undo is instant, and why the preview and the full-size export agree: both are drawn from the same description, one at screen size and one from your original pixels.
Large exports are prepared in a background worker where the browser supports it, so the page stays responsive. On photos larger than 3,072 pixels the soft edge is refined again at full resolution.
What is kept in your browser
The On-device button in the editor lists each of these with its size and clears any of them, or all at once. Nothing in that list ever leaves the device.
- The detection models, in the browser's cache, until you clear them.
- An unfinished edit, for up to 24 hours, so you can continue after closing the tab.
- Preferences: recent colors, favorite backdrops, your presets, brand colors and the model you last chose on each page.
Checking it yourself
You do not have to take our word for it. Open your browser's developer tools, choose the Network tab and clear it, then open a photo, cut it out and export it. You will see the site's own scripts and, on the first run, the model download. You will not see any request that sends your picture, its file name or anything measured from it. The privacy page describes exactly what, if anything, is measured about visits to the site.
Questions people ask
Can I check for myself that my photo is not uploaded?
Yes. Open your browser's developer tools, choose the Network tab, then open and export a photo. You will see the page's own files and, the first time, the model download, but no request that carries your picture.
Why does the first cut-out take longer?
The detection model is downloaded once (26 MB for people, 84 MB for objects, after a 0.2 MB face check in Auto) and kept in your browser. Later cut-outs start straight away.
Which tools work without any model?
Removing white or black, logos, signatures, the background eraser and adding a background to a transparent picture all run without a model.
Where are my unfinished edits kept?
In this browser's own storage for up to 24 hours, so you can continue after closing the tab. The On-device button in the editor lists everything kept and clears any of it.