How the background actually gets removed

A model, not a colour picker

The old way to drop a background was to pick a colour and delete everything close to it. That works on a studio shot against a plain wall and falls apart everywhere else, because a photograph of a person in a room has no single background colour to name.

This tool asks a different question. A neural network called U²-Net was trained on images paired with hand-drawn outlines of whatever the subject was, and what it learned is which pixels a person would say the picture is *of*. It does not look for a colour. It looks for the thing you photographed.

What it returns is not a yes-or-no answer per pixel but a confidence between the two, and that turns out to matter more than it sounds — it is the difference between an edge cut with scissors and one that blends.

Why your photo is never resized

The model only accepts one input size: 320 by 320 pixels. That sounds like a problem for a 24-megapixel photograph, and it would be if the picture were what came out the other end.

It is not. What comes out is the mask — a greyscale map of where the subject is. Your photo is shrunk to 320 square, the model looks at that, and the small mask it produces is scaled back up to the full size of the original and used as the alpha channel. The pixels you downloaded are the pixels you uploaded; only the mask was ever resampled.

Two useful things follow. A huge photo costs exactly the same to process as a small one, because the expensive part is a fixed size either way. And the mask is scaled up smoothly rather than as hard squares, which is what keeps the edge soft instead of stepped.

Three things to do with the space

Removing the background leaves a hole, and a hole is only sometimes what you want. The tool offers three answers and they produce genuinely different files.

Transparent

A PNG with a real alpha channel. Drop it onto a coloured slide or another photo and the background is simply not there. This is the default.

Solid colour

A PNG with the subject sitting on a colour you pick, and no transparency at all. What a marketplace listing usually wants.

JPG

The same flattened result as a JPEG. JPEG cannot store transparency under any circumstances, so the fill is not a style choice here — without it every cut-away pixel would come out black.

It runs here, not somewhere else

There is no API key in this product and no request to a third-party background-removal service. The model weights ship inside the worker container and inference happens on the same machine that received your upload.

The weights are U²-Net, published under the Apache 2.0 licence. That choice was deliberate: the best-scoring open model for this job is BRIA RMBG, and its licence forbids commercial use. Shipping it would have meant either breaking that licence or quietly limiting what you are allowed to do with your own picture.

Everything else about the file works the way it does for every other tool here — it is deleted two hours after it arrives, along with the result.

Where it struggles

Hair and fur against a busy background are the hardest case in this entire field, and this is the small version of the model. The edge is soft, which helps more than it sounds, but fine strands will not all survive and you should expect to touch some of them up.

A picture with no clear subject — a landscape, a texture, a pattern — has nothing for the model to isolate. Rather than hand back a blank image, the tool returns the picture intact. You have lost nothing except the time it took.

Frequently asked questions

Is the output really transparent?

Yes. It is a PNG whose header declares a colour type with an alpha channel, and whose background pixels have an alpha of zero — not a white rectangle that looks transparent on a white page.

Does it make my image smaller?

No. The model works at a fixed small size internally, but only the mask is scaled back up. The dimensions of what you download match the dimensions of what you uploaded.

How long does it take?

Under a second of actual inference for a typical photo, plus the time to upload the file and hand it through the queue. It is a heavy tool, so it is limited to one image at a time.

Can I put my own picture behind the subject?

Not yet. Today the choices are transparent, a solid colour, or a flattened JPG. If you want a custom backdrop, download the transparent PNG and drop it over your own image in any editor.