Background Remover
Remove the background from any photo without uploading it. The page runs a portrait-matting neural network (MODNet) in your browser, cuts out the subject with softly feathered edges, and hands you a transparent PNG at full resolution — nothing ever leaves your device.
Heads up: background removal runs a neural network in your browser. The model downloads once from this site (about 25 MB) and is cached afterwards — the first run needs a decent connection, and large photos take longer on modest devices.
Drop a photo
or drop anywhere on this page · Choose a file · 100 MB
Accepted: JPG, JPEG, PNG, WebP, AVIF, GIF, BMP, HEIC, HEIF
How does Background Remover work?
Automatic background removal is an image-matting problem: for every pixel, decide how much of it belongs to the subject and how much to the background. This page runs MODNet, a neural network designed for exactly this, compiled to run in your browser via onnxruntime-web. It predicts a soft alpha mask — not a hard cutout — so hair and slightly blurred edges keep their natural semi-transparency instead of turning into a jagged stamp.
Why a portrait model
MODNet was trained on photographs of people, which is also where background removal is wanted most: profile pictures, product shots, thumbnails. On portraits it is genuinely good; on pets, cars and objects it is merely decent, and the page says so up front. A general segmentation model would be slower and worse at the hair-level detail that makes a cutout look professional rather than pasted.
From mask to PNG
The network sees a downscaled copy of your photo and returns a mask, which the page scales back up to full resolution, softens with a light feather pass to remove stair-step edges, and writes into the alpha channel of your original pixels. Nothing is re-compressed or re-graded: the subject's colours are your photo's colours, pixel for pixel, and the download is a lossless PNG with real transparency.
The honest limits
Matting is estimation, and it fails gracefully rather than magically: wispy hair against a busy background, translucent objects, and subjects that touch the frame edge all confuse it. If the cutout looks wrong, a simpler photo — subject clearly separated from the background, decent light — fixes more than any setting could. There are no settings here on purpose: the model has no useful knobs, and sliders that do nothing are worse than none.
Example input
portrait.jpg — a photo of a person against any background
Example output
Transparent PNG of the subject at full resolution, ready to download
What options and edge cases does Background Remover support?
| Parameter | Type | Default | Behaviour & edge cases |
|---|---|---|---|
| Model | MODNet (neural network) | Automatic | A portrait-matting network running locally via onnxruntime-web. Downloads once (~25 MB) and is cached in the browser afterwards. |
| Edge treatment | automatic | Soft feather | The mask keeps the model's soft alpha; a light feather pass removes stair-step edges. No hard cutout, no halos. |
| Output | PNG | Full resolution | Lossless transparent PNG at your photo's original dimensions — the subject's pixels are untouched, only the background becomes transparent. |
| Input cap | 25 MP | — | Photos over 25 megapixels are declined instead of freezing your tab; the network itself sees a downscaled copy, so giant files gain nothing. |
| Privacy | local only | — | Inference runs in this tab. The only network traffic is the one-time model download; your photos are never uploaded. |
Frequently asked questions
Is my photo uploaded anywhere?
No. The neural network downloads once from this site (about 25 MB, cached in your browser afterwards) and then runs entirely in this tab via onnxruntime-web. Open DevTools on the Network panel while removing a background: no image data is ever sent anywhere.
What photos work best?
Photos of people — portraits, profile pictures, full-body shots. The model was trained on human matting, so hair and clothing edges come out clean. Pets, products and objects get a decent cutout but not the same hair-level precision. A subject clearly separated from the background beats any amount of re-running.
What format is the download?
A transparent PNG at your photo's original resolution. PNG is the right container here: it stores real alpha transparency losslessly, which is exactly what you want for compositing onto new backgrounds, slides or thumbnails. JPEG cannot hold transparency, so it is not offered.
Why is the first run slow?
The first run downloads the ~25 MB model, which dominates the wait on a slow connection. Afterwards it is cached in the browser and runs stay local — inference itself is a few seconds on a modern laptop, longer on modest phones. The progress bar tells you which phase you are in.
Can I remove the background from non-people photos?
You can try — the tool will attempt any photo. But be honest with yourself about the result: this particular model is a portrait specialist, so a car or a chair gets an approximate cutout. For product photography without people, a general segmentation tool will serve you better, and this page says that instead of pretending otherwise.
How is this different from remove.bg?
remove.bg uploads your photo to their servers and runs a larger model — it is better at hard cases. This tool's advantage is the opposite: your photo never leaves your device, there is no account, no per-image pricing and no watermark. For portraits on reasonable backgrounds the quality gap is small; for tricky commercial work, the server tools still win.