AI tools · Portraits → PNG

AI portrait background remover

Remove the background from a portrait with an AI model in your browser. Download a transparent PNG without uploading the photo.

Before you start

Static JPEG, PNG or WebP; 10 MiB, 20 megapixels. Output is resized to at most 1,600 pixels on its longest side. Designed for people; not a general product, pet or scene cutout tool. Fine hair, transparent objects and difficult lighting can fail.

First use downloads the model plus the inference runtime. Allow extra time and data. A recent desktop browser is recommended; older phones and restricted networks may fail.

Model weights: 7 MB · Xenova/modnet

AI processing runs on your device. When you load a model, your browser contacts jsDelivr and Hugging Face to download software and model files. Those services receive connection information such as your IP address; your selected files and text are not sent to them. Model files may stay in browser cache. Clear site data to remove the cache.

No file selected

Load the model when you are ready.

How this AI tool works

MODNet predicts a foreground alpha matte. The mask is applied to the resized portrait and exported as PNG with transparency. This tool uses the quantized model and keeps the photo on your device.

TDS · 2026-09-27

Xenova/modnet · fa2fa546052fba4c08921230a26cc69a333fca12 · q8 · Apache-2.0

Model and method sources

Questions about this tool

Can I remove a product background?

This model is trained for portrait matting. Product-only photos, animals and complex scenes may produce poor masks. Use it for people and inspect the exported edges.

What resolution do I get?

The longest side is capped at 1,600 pixels, without enlarging smaller inputs. The output is a transparent PNG.

All tools

Useful to someone else?

Share this page with your team. The link opens the tool or guide, with an empty workspace.

Link to this tool from your website

Select and copy this HTML into a relevant resource page or README. It links to an empty workspace; no files or results are shared. Attribution is optional.