Portrait background remover
Remove a person’s background locally and download a transparent PNG.
What Portrait background remover helps you do
Create a transparent cutout of a person without uploading their photo. The MODNet model estimates a soft foreground mask in your browser. It is designed for human portraits; inspect hair, hands and clothing edges before using the result.
How to use it
- Choose one still JPG, PNG or WebP portrait, up to 5 MiB and 8 megapixels.
- Read the local model download and performance information. Choose Remove background to load the engine only when you need it.
- Compare the original with the checkerboard transparency preview, then download the PNG. Cancel stops the worker; Start over clears the selected photo and result.
Input and output example
A 1,200 × 1,600 JPG portrait of one person against a plain wall.OUTPUT
A 1,200 × 1,600 PNG whose background is transparent. Fine hair may need manual cleanup in an image editor.
Supported formats and limits
- One still JPG, PNG or WebP at a time; up to 5 MiB (5,242,880 bytes), 8 megapixels and 12,000 pixels per side. Animated images, PDF files and HEIC are not supported.
- This is a human portrait model. Products, animals, multiple overlapping people, busy backgrounds, transparent objects and fine hair can produce incomplete cutouts. No manual brush refinement is included.
- The first run downloads the model and WebAssembly engine from this site. Your browser may retain these public runtime files in its HTTP cache. Photos and output are not uploaded or added to that cache.
- Inference uses a dedicated worker and one CPU thread. It may be slow or fail on older phones or browsers with limited memory. WebAssembly, canvas and a secure context are required.
- PNG export preserves pixel dimensions and original transparency but does not preserve camera, location or other source metadata. PNG output can be larger than the original JPEG.
How the result is calculated
RGB pixels are normalized as (channel / 255 − 0.5) / 0.5 for the MODNet fp32 model. Processing approximately preserves the image aspect ratio, targets a 512-pixel short edge, caps the long edge at 1,024 pixels for memory safety and rounds processing dimensions down to multiples of 32, with a 32-pixel minimum. Very wide or tall images can lose detail. The soft mask is resized with canvas interpolation to the original dimensions and multiplied by the original alpha channel. The output is a PNG with 8-bit transparency; no quality or accuracy percentage is promised.
Frequently asked questions
Is my portrait sent to an AI service?
No. This tool fetches public engine and model files from this site, then processes the photo on your device. There is no API key, account or per-image AI charge.
Will it remove the background from a product photo?
It may fail because MODNet is a portrait matting model trained for people. Use this page for portraits and check the preview. It is not a general object segmentation service.
Why does the first run take longer?
The local model and WebAssembly engine must be downloaded and initialized. Later runs may reuse your browser’s ordinary HTTP cache. Low-memory phones can still be unable to complete a run.
Can I cancel without waiting for the model?
Yes. Cancel terminates the dedicated worker and releases the tool’s active processing references. You can select another image or start again.