Runs on our server
Remove background
Cut the subject out of a photograph. This is a small segmentation model on our own machine — good on a clear subject against a plain background, and less good at hair, which the examples below show rather than hide.
Drop a photo here
Or press to choose one. PNG, JPEG or WebP.
What the model is, and why it is the small one
U^2-Net is a salient object detection network: it looks at a picture and decides which pixels belong to the thing the picture is of. We run u2netp, the lightweight version, at 4.6 MB of weights.
The full u2net is 176 MB and produces a slightly better mask, particularly at hair and fur. On a server with 7.7 GB of memory, no swap file and eleven other services — some of them medical — 171 MB of extra resident weights per request is not a rounding error. The small model was the right trade, and it is worth knowing which one you are getting rather than being told "AI".
Attribution: U^2-Net, Qin et al. 2020, Apache-2.0.
What it is good at, and what it is bad at
It is good at: a person, a product or an animal photographed against a background that is a different colour and reasonably plain. Product shots, headshots on a wall, an object on a desk. On those it is quick and the edge is clean.
It struggles with:
- Hair against a busy background. This is the classic failure and the small model is worse at it than the large one. Loose strands are lost or smeared.
- Glass, smoke, water and anything semi-transparent. The model produces a hard yes-or-no mask; a wine glass comes out as a solid shape.
- A subject the same colour as what is behind it. A grey cat on a grey sofa.
- Several subjects. It finds the salient object. Three people spread across a frame often gives you one of them.
- Motion blur, and very low contrast.
The before-and-after view sits on a checkerboard on purpose, so that transparency is visible rather than implied. A result that looks perfect on a white card can turn out to have kept a white halo.
Getting a better cutout
- Shoot or choose a photo where the subject is clearly separated from the background in colour and in focus.
- Crop first. The image is resized to 320 pixels square before the model sees it, so a subject occupying a tenth of a wide frame gets very few pixels of attention. Crop tightly and the mask improves noticeably — and cropping runs in your browser, free and instantly.
- Try the white or coloured background modes for a product listing. A cutout on a flat background is what most marketplaces ask for, and it hides small edge errors that transparency exposes.
- For difficult hair, expect to touch the result up in an image editor. This gets you 90% of the way in three seconds, not 100%.
The limits, and what happens to your photo
Up to 12 MB and 16 megapixels. If your photo is larger, run it through the image compressor first — that runs in your browser, takes a second, and costs nothing.
The output is always a PNG, because it is the format that carries an alpha channel and is read everywhere. It comes back at the original resolution: the mask is computed at 320 pixels and scaled up, which is why the edge is soft at very high resolutions.
One image is processed at a time and you are told your place in the queue. A cutout usually takes a few seconds.
Your photograph is uploaded, processed, and deleted as soon as the result exists. The result is deleted thirty minutes later, or ten minutes after you download it, whichever comes first, and there is a button on the result to delete it immediately.