How to use it
- Choose a photo with one or more people in it.
- Pick the Fast or Detailed model and select Remove background.
- Adjust threshold, softness and feather until the edge looks right; optionally choose a background colour.
- Compare before and after, then download the PNG.
How it works
The tool runs Google's open-source MediaPipe Image Segmenter in your browser. A small neural network looks at a 256 × 256 version of the photo and estimates, for every pixel, how likely it is to belong to a person. That mask is scaled up to your photo's size and used as the transparency of the PNG.
The Fast model (selfie segmenter, about 250 KB) separates person from background. The Detailed model (selfie multiclass, about 16 MB) recognises background, hair, skin and clothing; everything that is not background is kept. Both were trained for people, so they will not reliably cut out products, pets or cars.
Threshold sets how confident the model must be to keep a pixel, softness widens the transition between kept and removed, and feather blurs the edge by the chosen number of pixels. Because the mask is computed at low resolution, fine hair and thin gaps are approximated rather than traced exactly.
The model files and the WebAssembly runtime are fetched from public CDNs (Google Cloud Storage and jsDelivr) the first time and then cached by your browser. Results are saved as PNG with transparency, at up to 16 megapixels.
Examples
Profile picture on a brand colour
A head-and-shoulders selfie with the Fast model, threshold 50%, softness 20% and a replacement colour gives a clean avatar on a plain background.
Full-body cut-out for a flyer
A standing photo with the Detailed model, threshold 40% and 3 px feather produces a transparent PNG to place on a design.
Frequently asked questions
Is my photo uploaded?
No. The segmentation runs in your browser. Only the model and runtime files are downloaded from Google's and jsDelivr's servers; your photo never leaves your device.
Why does it not work for my product photo?
The models were trained to find people. Objects, animals and products are often not detected at all, or only partly.
Why are hair edges not perfect?
The model works on a 256 × 256 version of the photo, so very fine detail is estimated. Increase softness or feather for a more natural edge, or try the Detailed model.
Does it work offline?
After the first successful run the browser usually caches the model and runtime, but this is not guaranteed. The first use needs an internet connection.
Privacy
Your photo is decoded and segmented in your browser and the PNG is created on your device; the photo is not uploaded, stored or included in analytics. To run the model, your browser downloads the MediaPipe runtime from the jsDelivr CDN and the model file from Google Cloud Storage; those requests contain no image data. Only an anonymous event that the tool was used is counted, and your plan status is loaded from the site's account service.