Remove Background from Image (Free, No Watermark)

Drop in a photo and get it back with the background gone and the subject on transparency. Two modes are available: Fast, which returns a result in a few seconds and is right for most images, and HD, which runs a heavier model and takes noticeably longer but handles hair, fur and semi-transparent edges far better.

Both modes are BiRefNet, a bi-directional refinement segmentation network — Fast uses the lite variant, HD uses the full one. Rather than picking a colour to knock out, the model classifies every pixel as foreground or background and produces an alpha matte, which is why it works on a photo shot against a cluttered room and not just a studio sweep. The full model spends its extra time on edge refinement, and that is where the visible difference lies: soft boundaries, strands, and anything with partial transparency.

Where the result is essentially perfect: a clearly-lit subject against a background of a different tone, product shots on a plain surface, people photographed at a normal distance, animals, packaging, furniture. Where it takes real effort: flyaway hair against a busy background, fine mesh and lace, chain-link and wire, glass and bottles where you need what is behind them to show through, motion blur, and any subject whose colour closely matches what is behind it. In those cases run HD first, and accept that a fifteen-second wait is much cheaper than manual masking.

Some images no automatic tool will get right, and it is worth recognising them early. Anything where the boundary is genuinely ambiguous — a subject in front of a wall of the same shade, smoke, splashing water, out-of-focus foreground objects — needs a human decision about where the edge should be. Composite work destined for print, or a hero image that will be viewed large, is usually worth masking by hand in an editor. The honest rule: if the output will be seen at 200 px in a grid, automatic segmentation is fine; if it will be seen full-bleed, inspect it at 100% first.

Read the result at the edges, not in the middle. The recurring problems are a halo of background colour retained around the subject (most visible when you later place it on a background of the opposite tone), hair that has been cut into a smooth silhouette instead of individual strands, and thin features such as a chair leg or a strap that got dropped entirely. A quick way to check: place the cut-out on a solid colour very different from the original background — dark green against a white studio shot — and any fringing becomes obvious immediately.

Input can be PNG, JPG or WebP up to 20 MB. Output is a transparent PNG, or a JPG flattened onto white or black if you don’t need the alpha channel. PNG with transparency is the correct choice for anything you will composite later or place on a coloured page; JPG is far smaller and appropriate for a marketplace listing that requires a white background anyway. If you processed in Fast mode and the edges aren’t good enough, upgrading to HD reprocesses the same image without re-uploading it.

Practical scenarios this covers well: e-commerce listings where the marketplace demands a uniform white background, product images that need to sit on a coloured section of a landing page, headshots for a team page or a speaker profile, packshots for a slide deck, and cutting objects out for a collage, thumbnail or advert. It also removes distracting backgrounds from photos taken in whatever room happened to be available, which is most photos taken by anyone who does not own a studio.

On processing and privacy: segmentation runs on our server rather than in your browser, because the models are too large to load into a tab. Your image is sent over HTTPS, held in memory while the model runs, streamed straight back to you, and never written to disk or retained — it is not stored, not logged, and not used for training. There is no account and no watermark. Rate limiting applies to keep the queue responsive, so a burst of uploads in quick succession may be asked to wait.