Background Remover

About Background Remover

The Background Remover erases the background from a photo of a person and leaves you a clean cutout — a transparent PNG, or the same subject over any colour you pick. Unlike almost every other background remover on the web, it does not upload your image: the MODNet matting model is downloaded to your browser once, cached, and then run entirely on your own device, so private photos never touch a server. It is tuned for people rather than products, and it is built for the awkward part of the job — hair. The model produces a soft alpha matte instead of a hard cut-out mask, so edges blend into whatever you place behind them.

Key features

• Automatic portrait matting with soft, natural hair edges — an alpha matte, not a hard binary mask
• Transparent output, or flatten the subject onto any solid colour you choose
• Export as PNG, WebP, or JPEG; JPEG appears only once a solid background is set, since it has no alpha channel
• Runs on WebGPU where the browser supports it, and falls back to a quantised WASM build everywhere else
• Inference runs in a Web Worker, so the page stays responsive while an image is processed
• Each cutout is cached per file, so changing the background colour re-composites instantly instead of re-running the model
• Batch several photos at once, with a download progress bar for the one-off model fetch
• Accepts JPG, PNG, and WebP input, up to 30 MB per file
• 100% private — the ~7 MB model comes to your image; your image never goes anywhere

How to use it

1. Drop in one or more JPG, PNG, or WebP photos of people, or click to browse.
2. On the first run, wait for the model to download — a progress bar shows how far along it is. It is cached for every later visit.
3. Leave the background transparent, or pick a solid colour to place the subject on.
4. Choose the output format: PNG or WebP to keep transparency, JPEG once you have set a solid colour.
5. Save each cutout, or use Download all.

Tips & common mistakes

• The model is trained on photographs of people, so portraits, headshots, and profile pictures come out cleanly. Products, objects, pets, and logos are outside what it was built for and will not cut out reliably.
• The first run downloads about 7 MB of model weights. After that each image takes roughly a second, and your next visit starts instantly from the browser cache.
• Contrast is what the model has to work with: a subject that is similar in tone to the wall behind them produces a weaker matte than one shot against a clearly different background.
• Keep PNG or WebP if you need transparency. Exporting to JPEG flattens the alpha channel, which is why the tool only offers it once you have chosen a colour to flatten onto.
• Switching the background colour re-composites from the cached cutout — it does not re-run the model, so try several colours freely.
• A browser with WebGPU runs the full-precision model on the GPU; without it the tool uses a quantised CPU build, which is slower and very slightly softer at the edges. Both stay on your device.
• For a passport-style or profile photo, a plain solid background usually reads better than transparency — pick the colour here rather than compositing it later.
• The output needs a solid background rather than transparency? Convert it afterwards with JPG to PNG converter or the reverse.

Related tools