Portrait Background Remover
A transparent portrait, processed on your device.
Remove the background from a portrait locally with an AI person-matting model. Export a transparent PNG or WebP.
Your files
Files stay on your deviceDrop files herePick a file to start
PNG, JPEG, WebP (static) · up to 20 MiB per file
Limits & supported formats
Input: 1 file · PNG, JPEG, WebP (static) · up to 20 MiB.
Static images: up to 2,048 px per side. Animated inputs are not supported here.
Still image output: up to 2,048 px per side and 4,194,304 pixels. No animation is exported.
Detailed budgets: 40,000,000 decoded pixels, an estimated 192 MiB working allocation and 64 MiB per ZIP. These estimates help reject large jobs; they do not guarantee a browser process memory peak.
Add a file to get started.
Three steps. One moving picture.
01Choose a clear portrait
02Choose PNG or WebP for a transparent result
03Inspect hair and clothing edges, then download
Remove a portrait background
Separate a person from their background with MODNet. This model is trained for portraits; it is not a general object remover. Hair, small faces, busy backgrounds and colors similar to clothing can need manual cleanup. Keep the original and inspect the transparent result against a different background.
Choosing your settings
Choose a clear portrait up to 1024 × 1024 px. PNG and WebP preserve the estimated transparency; JPEG flattens it against your selected background. The model loads from this website only when you run the tool; your photo stays on your device.
If the result needs another pass
If the subject is partly missing or the background remains, try a clearer crop with one visible person. Do not use this portrait model to isolate products or animals; a solid-color source can instead be edited manually. AI alpha estimates are not an exact original mask.
Good to know
Can this remove any image background?
No. MODNet targets people in portraits. Product photos, animals, distant faces and crowded scenes may give poor masks.
Is the photo sent to an AI service?
No. Public model files load from the same website, then inference runs locally in a worker. Cancelling terminates the worker.