A cut-out that never leaves your device
Most “remove background” websites upload your photo to a server. This one does not. When you click the button, your browser downloads a self-hosted model from this site, runs it locally, and builds a transparent PNG. That is slower on the first run than a cloud API, and the cut-out is not as polished as a paid, hair-aware service — but the picture stays on your computer, which matters for a passport photo, a product shot or a personal portrait.
How to remove a background
- Select a photo. Click Select a photo and choose a JPG, PNG, WebP or AVIF. The picture is opened in your browser; nothing is uploaded to a server.
- Remove the background. Click Remove background. The first time in a tab, the page downloads an on-device model (about 42 MB) plus a WebAssembly engine (about 15 MB) from this site, not from a third-party CDN. Later runs reuse the copy already in memory.
- Adjust the edges if you need to. After the model has produced a low-resolution mask, this page sharpens it with a guided filter, zeros very faint leftover alpha and lightly median-filters only the remaining low-alpha fringe, and, by default, defringes colour that leaked from a light background into hair and other semi-transparent pixels. Edge softness and the Defringe toggle re-run only that edge pass, not the model or its download.
- Check the preview and download. The result is a transparent PNG, shown on a checkerboard. Download it if the cut-out looks right. Fine hair, fences and busy scenes can still be messy — try a clearer photo of one subject if they are. The refine pass helps a light fringe; it does not invent missing strands.
What the model is, and what it gets wrong
The model is Silueta, a reduced-size version of U²-Net (a published subject-segmentation network) at about 42 MB, plus a 15 MB ONNX Runtime engine. Both are stored on this site and loaded only when you use this tool, so a visit to the image compressor or a PDF page does not pay that cost. It was trained for general objects and people, not specifically for wispy hair or transparent objects. After that mask is produced, a guided filter (He, Kaiming et al., 2013) sharpens the edge against the original photo, a light cleanup clears faint leftover alpha, and Defringe — on by default — decontaminates colour that leaked from a light background into semi-transparent pixels. Expect a cleaner edge on a mug, a book or a person against a reasonably plain wall. Fine hair, a bicycle wheel, a fence, or hair against a busy room can still leave holes or leftover background; the refine pass reduces a pale fringe, but it does not invent missing strands or match a paid, hair-aware service.
If the result is close but not perfect, you can tidy the PNG in any editor, or try a tighter crop first with the crop tool. To convert the PNG to another format, use the Image Converter or SVG to PNG for vector artwork (this remover is for photographs, not SVGs).
Related tools and guides
- Crop, rotate and flip: trim the subject first
- Image Converter: save the PNG as JPG or WebP
- PNG to JPG: flatten onto a colour if you need a JPG
- Signature maker: a transparent PNG of a signature
- SVG to PNG
- Image to text: read text in the photo instead
Background remover FAQ
Is my photo uploaded?
No. The image and the model both run in your browser. The only files fetched are the model and the engine, and those come from this site when you click Remove background — not from a third-party CDN, and not when you merely open the page.
What model is this?
Silueta, a reduced-size U²-Net model (about 42 MB) published with the rembg project and based on U²-Net (Xuebin Qin et al.), under the Apache License 2.0. It is a general-purpose cut-out, not a specialised hair or portrait service. After the 320×320 mask is produced, a guided filter (He et al., 2013) upsamples it against the photo, a light cleanup zeros dust and smooths only very low alpha so solid edges and stronger strands stay sharp, and Defringe — on by default — estimates the background colour near edges and removes it from semi-transparent pixels. Quality is usually good on a single clear subject. Fine hair, glass, fences or a crowd can still leave artefacts.
What does Edge softness do?
It only changes the guided-filter window used to upsample the mask. The default is 35, a little below the middle of the slider: enough to tighten a fringe without following every background texture. Zero is the raw bilinear upsample, with no guided filter — useful if you want to compare. Moving the slider does not re-download or re-run the model.
What does Defringe do?
On by default, it looks at colours next to the cut-out edge and pulls a light (or tinted) background out of semi-transparent pixels, which is what usually causes a pale halo around hair. Turn it off if a dark fringe appears instead, or if you prefer the unprocessed colours. Like Edge softness, it re-runs only the edge pass.
Why is the first run slow?
The model and the ONNX Runtime engine have to be downloaded from this site and initialised. On a typical broadband connection that is tens of seconds the first time, then much faster in the same tab. We do not preload them on other pages.
What do I get as a file?
A PNG with a transparent background, at the same size as the photo (or scaled down if a side was larger than 4096 pixels, so the tab stays usable).
Which pictures work best?
A single person, product, pet or object, reasonably sharp, with a background that is not full of competing shapes. A logo on a plain backdrop usually cuts cleanly. A group photo, a busy street, or hair against a similar colour often leaves artefacts.
Can I put a new background behind the subject?
This page only produces a transparent PNG. You can place that PNG on a slide, in a document, or in an editor. To drop it into a PDF page, use a PDF editor you trust, or start with Images to PDF if the subject should sit on its own page.