Free · from 4Brains Labs
Drop in a photograph and get the subject back on transparency, at the size you uploaded it. No account, no watermark, and no cap on what you download.
Drop an image herePick a photo to cut out
JPEG, PNG or WebP · up to 25 MB · no account, no watermark
Your image is sent to our own server in India, processed on a graphics card there, and held only in memory — it is gone as soon as the result reaches you. Nothing is written to disk, there is no history and no account. We will not pretend the picture never leaves your device: it does, once, and then it is deleted.
Upload a photograph and it comes back with the background gone: the subject on transparency, ready to drop onto a white listing page, a poster, a slide or another photograph. The work is done by a segmentation model called BiRefNet, running on a graphics card in our own office, and it takes a few seconds.
It is good at the ordinary cases — a person, a product, a pet, a piece of furniture, anything that is clearly the subject of the picture. It is not magic, and the section on where it struggles is there because knowing the failure cases in advance saves more time than any feature we could add.
The cut-out you download has exactly the pixel dimensions of the file you uploaded. A 24 megapixel camera photograph comes back at 24 megapixels. This is worth saying plainly because it is the thing most free background removers hold back.
remove.bg, the best known of them, caps a free download at about 0.25 megapixels, roughly 625 × 400 pixels. That is enough to check that the cut-out worked and not enough to print, to put in a listing, or to place in a layout at any size. Full resolution is what the credits buy. Slazzer, Removal.AI and Cutout.Pro use much the same cap; Photoroom limits web exports to 2000 × 2000.
| What you upload | You get here | A 0.25 MP free download |
|---|---|---|
| Phone photo, 4000 × 3000 (12 MP) | 4000 × 3000 | 577 × 433 |
| Camera photo, 6000 × 4000 (24 MP) | 6000 × 4000 | 612 × 408 |
| Square product shot, 2000 × 2000 (4 MP) | 2000 × 2000 | 500 × 500 |
The right-hand column is what a 0.25 megapixel cap works out to at each shape, and describes the free tier of those services as it stood in September 2026. Their paid plans return the full size, the same as we do.
One honest qualification. The model itself looks at your picture at 1024 × 1024 and the mask it produces is then scaled back up to your dimensions, so the edge on a very large image is smooth rather than pixel-exact. Every pixel of the subject is your own, at your own resolution; it is the boundary between subject and background that is interpolated. In practice that is invisible on a normal photograph and occasionally visible on a picture with a very intricate outline.
Automatic segmentation is very good and not perfect. These are the cases where it commonly gets things wrong. None of them are specific to this tool: they are hard for every background remover, paid ones included.
Fine hair and fur
Individual strands are thinner than the mask can resolve, so flyaway hair is softened or dropped, and a halo of the old background can survive between strands. Hair against a plain, contrasting background works far better than hair against foliage or a crowd.
Glass, water and anything see-through
The model decides what is subject and what is not. Something genuinely transparent — a wine glass, a bottle, a veil, smoke, a car windscreen — is neither, so it usually comes back opaque, keeping whatever was behind it, or is cut away entirely.
Thin structures
Wires, cables, bicycle spokes, chair legs, railings, jewellery chains, spectacle arms and plant stems. Because the mask is computed at 1024 pixels across, a feature narrower than about four pixels in a 4000 pixel-wide photograph can simply vanish.
Low contrast
A white shirt against a white wall, a black dog on a dark sofa, grey machinery in a grey workshop. If you have to look twice to find the edge, so does the model, and it will place it in the wrong spot.
Motion blur and shallow depth of field
Where the boundary itself is a gradient over twenty pixels there is no correct answer, only a choice. The mask tends to cut inside the blur, which can make a moving hand or a defocused shoulder look clipped.
More than one subject
The model looks for the salient subject. In a group photo it usually keeps everyone, but a person half out of frame at the edge, or an object the photograph does not emphasise, may be removed along with the background.
Shadows and reflections
A shadow cast on the floor is part of the background and goes with it, which can leave the subject looking as though it is floating. Reflections in a mirror or a polished table are treated inconsistently.
Heavily compressed pictures
A photo that has been through a messaging app several times has blocky artefacts along every edge. The model sees those blocks as real detail and the cut-out gets a ragged outline. Use the original file if you still have it.
If a result is wrong, it will usually be wrong in one of these ways, and a different photo of the same thing often fixes it. Nothing is stored, so trying again costs nothing.
The download is a PNG with a real alpha channel: the background is not white, it is absent. That is what you want when the cut-out is going into something else, because it will sit on whatever colour, texture or photograph is behind it without a rectangle around it.
Choose a white background instead when the destination demands one. Marketplace listings are the common case: Amazon, for instance, requires the main product image to be on pure white, and a transparent PNG uploaded there can be rendered against grey or black. Print workflows and older office software are the same. Some of them fill transparency with black rather than white, which is a surprise best avoided at the last step rather than discovered on a proof.
On formats: PNG is lossless and carries transparency, which is why the result comes back as one, and why the file is larger than the JPEG you uploaded. JPEG cannot store transparency at all. If you need a JPEG in the end — most listing pages and email systems prefer one — composite the cut-out onto its background first, then export to JPEG. Converting the transparent PNG directly usually gives you a black background. WebP sits in between: it supports transparency and is smaller than PNG, and is a reasonable choice for the web once you are finished editing.
This is a server-side tool and we would rather be exact about that than vague. Your picture is uploaded to our server, passed to a graphics card we own and run ourselves in our own office, and the cut-out is sent straight back to your browser. It is never sent to a third-party service.
Nothing is written to disk. The image exists in memory for the seconds the work takes and is released when the response has been sent. There is no gallery, no history, no account and no copy kept for training, review or anything else. If you need to delete something afterwards, there is nothing to delete. The full detail is in the privacy policy.
Yes. There is no sign-up, no credit, no watermark and no cap on the size of what you download. The GPU is one we already run for other work, so an extra image costs us very little.
JPEG, PNG and WebP. The file has to actually be one of those: we check the bytes rather than trusting the file name. You always get a PNG back, because PNG is the common format that can carry transparency.
The file has to be under 25 MB. That is a limit on the file, not on the resolution: a 25 MB JPEG can easily be 50 megapixels, and it comes back at 50 megapixels. Almost every phone and camera photograph is well under the limit, so if you hit it you are probably holding a scan or an export at print settings.
Ten a minute and a hundred an hour. There is one graphics card behind this, and the limit is there so that one visitor cannot occupy it for everybody else. If you need more than that, write to us rather than working around it.
Yes. The model looks for the main subject of a photograph, whatever that is, so products, furniture, food, animals, cars and plants all work. It is least certain when a picture has no single obvious subject.
Yes. The picture is yours and stays yours; we claim nothing over it and add no watermark. The model behind the tool, BiRefNet, is MIT licensed, so nothing in the pipeline restricts what you do with the output.
There is one graphics card, and it ran out of memory part-way through your image, usually because another request was already on it. Nothing was lost. Wait a few seconds and send the same file again.
No. The image is held in memory for the few seconds it takes to process, and is gone once the result has been sent to you. There is no history, no gallery and no account for anything to be attached to.
Something behaving oddly, or a photo it handles badly that you think it should not? Send it to contact@4brains.in.