How it works
- You add a photograph. Nothing happens on the network yet — the model is not downloaded until you press the button, because it is large and most visitors are only reading the page.
- When you press the button, the segmentation weights and the inference runtime download — about 95 MB together, measured on a real run. That download has its own progress state in the interface, separate from processing, so a slow connection reads as a slow download rather than a hung tool. Whether a second run skips it depends on your browser holding those files in its HTTP cache; the library does not manage a cache of its own.
- The model runs in a Web Worker via WebAssembly, with WebGPU used where the browser exposes it. It produces an alpha matte — a per-pixel opacity map — rather than a hard cut-out, which is what allows semi-transparent edges instead of a jagged outline.
- The matte is composited with the original pixels onto a canvas with a transparent background, and the result is encoded as PNG. JPEG is not offered because JPEG has no alpha channel; a JPEG 'cut-out' is just a cut-out on white.
- The output is a blob URL held only by this tab. The photograph never leaves the device — this is a case where an on-device model is doing work that most competing tools send to a GPU server.
Limitations
Hair and thin edges are where it fails, and it fails predictably. Flyaway strands, fur, mesh fabric, chain-link, glass and motion blur all produce boundaries where the matte guesses. A portrait against a plain wall with tied-back hair comes out clean; the same person outdoors on a windy day against a busy background will need manual work on the edges.
The first run downloads about 95 MB. On a metered mobile connection that is a real cost, and it is the reason nothing downloads until you explicitly start. If the connection drops partway, the download restarts rather than resuming. A smaller quantised model exists at roughly half the size, and it is not used here because quantisation damages exactly the thin edges this tool already handles worst.
Output resolution is capped by what the browser canvas and your device memory can hold. Very large images are processed at a reduced working resolution and the matte is scaled back up, which softens the edge slightly on images beyond about twenty megapixels.
The model has a strong idea of what a subject is — people, animals, products, vehicles. Photographs where the subject is ambiguous, such as an architectural detail or an abstract texture, will produce a confident and wrong result rather than an uncertain one.
Questions
Why is there such a large download?
About 95 MB, because both the model weights and the inference runtime have to reach your device — that is the price of not sending the photograph to a server. Nothing downloads until you press the button, and if your browser keeps the files in its cache a later run will start faster.
Is my photo uploaded?
No. The model comes to the photo rather than the photo going to the model. This is the tool where that distinction is worth the most — most background removers upload your image to a GPU server, and portraits and product shots are exactly the images people would rather not hand over.
Why are the edges of the hair wrong?
Because hair is the hardest case in image matting. The model estimates a per-pixel opacity, and thin strands against a detailed background have no clean answer. Simple backgrounds, even lighting and hair that is not moving all improve the result substantially.
Can I download it as a JPG?
No, and that is not an omission. JPEG has no alpha channel, so a JPEG version of a cut-out would have a solid background — which is the thing you removed. PNG is the format that can carry transparency.
Do I need an account?
No. There is no sign-up, no email step and no usage counter. The tool is paid for by the ads on this page.
When people use it
Most people arrive with a product photograph and a marketplace that wants a white or transparent background. The listing gets rejected, or the photo sits awkwardly against the page, and the fix is a cut-out rather than a reshoot. For a flat-lay product on a plain surface, the model handles this in one pass.
The second group wants a person on a different background — a profile picture, a team page, a slide. Here the honest advice is to look at the edges before you use it. A head-and-shoulders shot with tidy hair is usually finished; anything else is a good starting point that saves you the outline but not the touch-up.
Cut-outs frequently end up in a document rather than on a web page — a product sheet, a lookbook, a set of listings a buyer wants as one file. Images to PDF assembles them into a single PDF in the order you set, one image per page.
How it compares
Against the well-known online removers, the trade is explicit: you spend about 95 MB of download and get a tool that never sees your photograph, with no credit system and no watermark on the output. Against a desktop editor, Photoshop's select-and-mask with a refine-edge brush still beats any automatic matte on hair, and if the image is going into print or a paid campaign, that manual pass is worth the time.