Background Remover

What this tool does

You drop in an image — JPG, PNG or WebP — and an AI model separates the subject from the background. You can download the result as a PNG with transparency or swap the background for white, black, or a color of your choice.

What sets it apart from most online background removers is where the work happens: everything is processed inside your browser. The image is never uploaded to any server. No account, no daily image limit, no watermark, and no copy of your photo on somebody else’s hard drive.

How to use it

1. Drag the image in, or click to choose one.
2. Press Remove background. The first time, the AI model is downloaded; after that it stays cached and subsequent images are near-instant.
3. Choose the background for the result: transparent, white, black, or a custom color.
4. Download the PNG.

The grey checkerboard in the preview is not part of the image: it is the standard way of showing transparent areas.

Why in-browser processing matters

A conventional background remover works by uploading your file to a server, which processes it and sends the result back. That has three consequences that rarely get spelled out:

• The image leaves your machine. For a product photo that hardly matters; for a client document, a screenshot with internal data, or a photograph of a person, it matters a great deal.
• There is a quota. Server cost is real, so the business model almost always ends in a monthly free-image limit, reduced resolution, or a watermark.
• It depends on the service continuing to exist. Here, once the model is downloaded, the tool works even with no connection.

The price you pay is that your machine does the work: the first run takes longer because the model has to come down, and on a modest computer the inference is not instant.

Two engines, and why there are two

The tool picks automatically between two models depending on what your browser supports:

• BiRefNet lite, if your machine has WebGPU available. This is the good one: far more accurate on difficult edges, and it runs on the graphics card.
• U²-Netp, if not. A lightweight model (around 4.4 MB, Apache-2.0 licensed) that runs on the CPU. It gives a reasonable cut-out and works practically everywhere.

You do not have to choose: the page detects support and uses the better of the two it can actually run. If the result looks poor, the most likely cause is that you are on a browser without WebGPU — typically a phone or an older browser — and are getting the fallback model. A recent desktop browser changes the result visibly.

What it gets right and where it fails

No automatic cut-out is perfect. Knowing in advance where it breaks saves time:

It works very well on people and portraits, products on a plain background, animals, objects with a defined silhouette, and generally any image where the subject contrasts with what is behind it.

It struggles with:

• Loose hair and fur. The classic hard case. BiRefNet handles it fairly well; the fallback model leaves hard edges.
• Glass, smoke, veils and anything semi-transparent. The model decides per pixel whether it is subject or background, and semi-transparent is neither.
• Subject and background in the same color. With no contrast there is no edge to detect.
• Several candidate subjects. The model looks for the main element; in a scene with three equally prominent objects it may keep just one.

One detail about the input image

Input resolution matters more than you would expect. Segmentation models work internally at a fixed size, so a huge image gets scaled down before analysis, and fine detail — strands of hair, threads, antennae — is lost in that step, not in the cut-out.

In practice: an image between 1,000 and 2,000 pixels on its long side usually gives a better edge than one at 6,000, and takes less time. If your original is very large, resizing it before running it through here is almost always a good idea.

After the cut-out

The result always comes out as PNG, because that is the format that preserves transparency. JPG cannot: it has no alpha channel, and saving a cut-out there turns everything transparent into white.

If the final destination is the web and you do not need transparency — because you are placing the subject on a solid color — converting the PNG to WebP cuts the weight considerably. And if you simply want it to weigh less while keeping the format, the image compressor does that job.