How it works

Run a U²-Net background-removal model in your browser to isolate the subject. Outputs a transparent PNG.

First use, the tool downloads a ~50 MB ML model (the @imgly/background-removal package, which wraps U²-Net via ONNX Runtime Web). After that, the model is cached in your browser, so subsequent runs are quick. The model analyses the image, identifies the subject, and produces a PNG with the background replaced by transparency. Works well on photos with a clear subject (people, products, animals). Struggles with low-contrast subjects, fine details like hair on a busy background, and translucent objects. For pro-quality results on tricky images, a human still beats the model, but for product shots, profile pictures, and casual cutouts, it's very good. Everything runs locally: the image and the model both stay in your browser, no upload, no server inference. Output filename has -no-bg.png appended.

How to use it

1. Drop your image. Photos with a clear subject — people, products, animals — work best.
2. Wait for the model on first use. A ~50 MB U²-Net model downloads once and caches; later runs take seconds.
3. Let it process locally. Inference runs in your browser via WebAssembly — the image never leaves your device.
4. Download the cutout. The subject downloads as a transparent -no-bg.png.

Frequently asked questions

Why is the first run slow?

The ML model is about 50 MB, downloaded from a CDN on first use. After that it's cached, so subsequent runs only run inference (a few seconds on a modern device).

Does this upload my image to a server?

No. The model runs entirely in your browser via WebAssembly. Both the image and the model are local.

What images does it struggle with?

Low-contrast subjects against busy backgrounds, fine hair detail, translucent objects (glass, smoke), and anything where the subject blends with the background.

What's the output format?

PNG with a transparent background. You can paste it onto any other image or use it directly on the web.