How it works

This tool runs BiRefNet, an image segmentation model, inside your browser with ONNX Runtime Web. For each photo the model works out, pixel by pixel, how likely it is to belong to the main subject. That estimate becomes the transparency of the output: subject pixels stay solid, background pixels become transparent, and the edges in between are partly see-through so hair and soft outlines blend into a new background.

The model and its engine are about 110 MB together. To keep the page light they are only downloaded when you press Remove background, the page tells you the size first, and your browser stores them afterwards. When your browser supports WebGPU the model runs on your graphics card; otherwise it runs on your processor with WebAssembly, using several cores where it can.

What to expect

Clear subjects such as a product on a table, a person in front of a wall or a car in a street usually come out cleanly.
Fine detail like flyaway hair, fur, lace or the gap between fingers is the hardest part. Edges may be a little soft or keep a trace of the old background.
Busy or low-contrast photos, where the subject is a similar colour to its surroundings, can lose parts of the subject or keep parts of the background.

The output is always a PNG, because JPG cannot store transparency. A PNG of a large photo can be several megabytes; to make it smaller for the web, Compress image can convert it to WebP with transparency kept. iPhone photos in HEIC format can be turned into JPG first with HEIC to JPG.

Privacy and licences

Your photos never leave your device. The model files are served by this website, checked against fixed checksums, and kept in your browser’s storage. BiRefNet is by Peng Zheng, Dehong Gao, Deng-Ping Fan, Li Liu, Jorma Laaksonen, Wanli Ouyang and Nicu Sebe, released under the MIT licence; the 512 × 512 ONNX export used here is by Studio Ludens, also MIT. ONNX Runtime is by Microsoft, MIT licence.

Frequently asked questions

Are my photos uploaded anywhere?

No. The AI model runs inside this browser tab, using your graphics card (WebGPU) when the browser allows it and your processor (WebAssembly) otherwise. The only download is the model itself, which comes from this website. Your photos are read from your disk and the PNGs are saved back to it.

Why does it download about 110 MB first?

That is the AI model (about 94 MB) plus the engine that runs it (14 to 16 MB, depending on whether it uses your graphics card). Nothing is downloaded until you press Remove background, and your browser keeps both in its storage, so later visits start without downloading them again. If you clear your browser data, they are downloaded again the next time.

How long does it take?

In our tests, about 2 seconds per photo on a desktop processor and under a second on a graphics card, after the model has loaded. Laptops can take a few seconds per photo. Browsers with WebGPU (current Chrome and Edge on most computers) use the graphics card. Phones and older computers are slower and may run out of memory with very large photos; if that happens, try one photo at a time.

Will the result have the same size as my photo?

Yes. The model looks at a 512 × 512 pixel version of your photo to decide what is subject and what is background, and that mask is then scaled up and applied to the original image. The PNG has the same width and height as the photo you added.

What kind of photos work best?

Photos with one clear subject: a person, an animal, a product or an object that stands out from what is behind it. Busy scenes, subjects that blend into the background, and very fine details such as loose hair or fur can come out rough at the edges. Group photos usually keep everyone in the foreground.

Which AI model is this, and can I use the results commercially?

It is BiRefNet lite, a segmentation model by Peng Zheng and co-authors, in a 512 × 512 ONNX export by Studio Ludens. Both the model and the export are released under the MIT licence, which allows commercial use. The images you make are yours.