How UNBG works
UNBG removes the background from a photo inside your web browser. The page downloads the app and an AI model, and your own device does the work. Here is what happens, step by step, and what does and doesn't leave your device.
1. You choose an image
When you drop or select a file, your browser reads it from your device. UNBG checks it before doing anything else:
- What the file really is. UNBG looks at the file's contents, not its name or extension, and accepts still JPG, PNG and WebP images.
- Whether it is safe to open. Empty, corrupted, oversized (over 48 MB) or extremely large (over 60 megapixels) images are rejected with a clear message rather than crashing the tab.
- Orientation. Photos from phones often store their rotation separately; UNBG applies it so the result faces the right way.
A preview appears immediately. At this point the image exists only in your browser's memory.
2. The AI model is prepared
UNBG uses BiRefNet lite 1024 (fp16, WebGPU), an open-source image-segmentation model released under the MIT licence (BiRefNet by Peng Zheng et al. (MIT, Copyright (c) 2024 ZhengPeng); ONNX conversion by jiabins0303.). Model details and attribution
- First use: your browser downloads the AI engine and model files (about 142 MB (about 113 MB in compatibility mode)) from unbg.me (the same site). These are ordinary downloads of fixed files; your image is not part of the request.
- Integrity check: each file is checked against a SHA-256 fingerprint published in UNBG's model manifest. If a file doesn't match, UNBG refuses to use it.
- Caching: UNBG tries to store the verified model in your browser's cache (Cache Storage), so later visits can reuse it. Updates, storage eviction or unavailable caching can require another download. You can clear it at any time: how to clear it.
3. The background is removed on your device
The model runs inside a background thread of your browser tab (a Web Worker), so the page stays responsive.
- WebGPU is used when your browser and graphics hardware support it. The model then runs on your GPU.
- WebAssembly is the fallback. The model runs on your processor instead, which can be slower.
- If neither is available, UNBG tells you so instead of failing silently.
The model looks at a resized copy of your photo (1024 px square) and produces a mask: a greyscale map of what is subject and what is background.
4. The full-size result is created
The mask is scaled back up to your photo's original dimensions and applied to your original pixels, not the resized copy, so detail is kept.
- If your device doesn't have enough memory for a very large image, UNBG says so and offers a clearly labelled smaller result. It never silently shrinks your image.
- Optional Refine controls let you adjust edge hardness, softness and colour fringe before you download.
5. You download a PNG
- Transparent keeps the background see-through (PNG with an alpha channel).
- White or Custom colour fills the background with a solid colour.
- The file is named after your original, for example
holiday-photo-unbg.png. - Camera metadata such as EXIF and GPS location is not copied into the result.
The PNG is created in your browser and saved by your browser's normal download mechanism.
6. Several images at once
If you choose more than one file (up to 50 at a time), UNBG queues them and processes them one after another. You can download each result or all of them as a single ZIP file, which is also built in your browser. More about batch processing
What leaves your device, and what doesn't
| Item | Sent over the network? | Details |
|---|---|---|
| Your image, its pixels and its filename | No | Processed in the browser tab only |
| The mask and the finished PNG or ZIP | No | Created in the browser and saved by your browser's download mechanism |
| Requests for the website's own files (HTML, CSS, JavaScript, icons) | Yes, to unbg.me | Standard page loading. No image data. |
| Requests for the AI engine and model files | Yes, to unbg.me (the same site) | First use, then reused when caching is available. Updates or cleared storage may require another download. No image data. |
| Analytics | No | UNBG does not run analytics. See the privacy policy. |
Before each release, an automated test processes an image containing a unique marker. It confirms that no request contains the marker and that nothing is uploaded.You can check it yourself too: here's how.
Browsers
[tested browser list: published from the QA browser matrix, to be confirmed]
If your browser isn't listed, try the latest Chrome, Edge, Firefox or Safari. UNBG checks what your browser supports when you start and tells you if it can't run.
Limits worth knowing
- Tricky subjects. Fine hair against a busy background, glass, smoke and very low-contrast photos are harder for any automatic tool. Check edges before you use a result somewhere important.
- One subject works best. Several overlapping people or objects may be kept or removed together.
- Memory. Very large images need a lot of memory. Phones and older computers may be offered a reduced-size result.