Background Remover
Cut out a subject and remove the background — AI in your browser. Batch, erase/restore brush, colour or photo behind, drop shadow, full-size PNG or JPG.
Drop images here or click to browse
Cut out the subject and remove the background — the AI runs entirely in your browser, your image is never uploaded.
PNG, JPG, WebP, GIF (first frame) or BMP, up to 40 MB each. Drop several at once to do them all in one go.
🔒 Everything runs on your device — your image is never uploaded. The fast model (U²-Netp) and its runtime come from this site, so it keeps working offline after the first load. The accurate model (BEN2) is too big to host here: its weights are fetched once from the Hugging Face model hub, then run locally on your GPU (WebGPU) or CPU in a background thread. If that fetch is blocked or fails, the fast model finishes the job and this page tells you why.
Remove an image background — privately, in your browser
Cut out the subject of a photo and delete the background, then download a transparent PNG — or put the subject on white, black, any colour you pick, or a photo of your own. Unlike most background removers, this one does not upload your image anywhere — the neural network runs directly on your own device.
A whole folder at a time. Drop several pictures at once (or paste a screenshot) and they are all cut out in one run. Each keeps its own name — photo.jpg is saved as photo-nobg.png — and you can save them one at a time or take the batch as a single ZIP. Anything dropped that is not a picture is named on screen instead of being quietly swallowed. PNG, JPG, WebP, GIF (first frame) and BMP are accepted, up to 40 MB per file.
Your own resolution, not the model's. The model works on a small square of its own — 320 pixels for the fast model, 1024 for the accurate one — but only its mask is kept: the mask is scaled up to the picture and laid over the original pixels. A 1200×900 photo comes back 1200×900, in PNG, in JPG and inside the ZIP.
Two models. Fast (U²-Netp) is served from this site — about 18 MB of model plus WebAssembly runtime, downloaded once and then reused from your browser's cache. Accurate (BEN2) is too large to host here, so its weights are fetched once from the Hugging Face model hub (219 MB) and then run on your own GPU through WebGPU, or on the CPU, in a background thread so the page never freezes. If that download is blocked or fails, the fast model finishes the job and the page tells you exactly why. Both work well on people, products and clearly-defined subjects; very fine detail like wispy hair comes out softer on the fast model than on a paid server-side tool. Refine edges, offered with the fast model, fills holes inside the subject, removes stray specks and feathers the edge.
Fixing what the model got wrong. Set the brush to Erase and drag on the picture to rub background away, or Restore to paint the subject back, with a size slider from 4 to 120. Strokes are painted into the mask at the picture's own resolution, so a correction is as sharp as the photo is, and within one stroke Erase can only ever make a pixel more see-through while Restore can only make it more solid — overlapping passes never fight each other. Undo my brushwork puts the model's own answer back.
Behind the subject. A background photo is scaled up until it fills the frame and centred, so it is cropped rather than squashed to the subject's shape. The Shadow slider (0–100%) drops the subject's own silhouette behind it, offset by 1.5% and blurred by 2% of the picture's long edge, so the same photo looks the same at any resolution. Save as PNG or JPG; because JPG has no alpha channel, choosing it over a transparent background is refused with an explanation rather than filling the hole with black behind your back.
Frequently asked questions
Is my image uploaded to a server?
No — your image never leaves your device, on either model. The fast model (U²-Netp) and the WebAssembly runtime are served from this site. The accurate model (BEN2) is 219 MB, too big to host here, so its weights are downloaded once from the Hugging Face model hub — that is a download of the model to you, not an upload of your photo — and it then runs locally in a background thread.
Why is the first run slow?
Your browser downloads the model you picked and keeps it: about 18 MB for the fast model (its ONNX weights plus the WebAssembly runtime, both served from this site), 219 MB for the accurate one from the Hugging Face hub. That happens once — later runs reuse the cached copy, and changing the background, the shadow or the brush never runs the model again.
What happens if the accurate model cannot be downloaded?
The fast model finishes the job instead, and the page says exactly why — you are offline, huggingface.co did not answer (it may be down, or blocked on your network), the download stalled, or the model would not load on this device. You always get a result rather than an error, and a "Try the accurate model again" button re-checks whenever you want.
The edges around hair look soft — why?
The fast model is deliberately lightweight so it can run for free on your device. It handles most subjects well, but very fine detail is softer than a heavy server-side service. Three things help: switch to the accurate model, leave Refine edges on, or fix the few wrong pixels by hand — set the brush to Erase or Restore and drag on the picture. The brush paints at the photo's own resolution, and "Undo my brushwork" puts the model's answer back if you overdo it.
Can I keep a white background instead of transparency?
Yes, and more than white. Background offers Transparent, White, Black, a colour picker, and "Use a photo" — your own picture behind the subject, scaled up until it fills the frame and centred so it is never squashed. Add a drop shadow with the Shadow slider if the subject needs to sit on the new background rather than float over it.
Can I do several images at once?
Yes. Drop or paste as many as you like in one go; they are listed by name, cut out in a single run, and a tick marks each one as it finishes. Save them one at a time, or click "Download all as ZIP" for the whole batch — each file inside is named after its own picture. If the accurate model turns out to be unreachable, that is discovered once for the batch rather than once per picture.
Does it shrink my image?
No. The model looks at a small square (320 px for the fast model, 1024 for the accurate one), but only the mask it produces is used — that mask is scaled up to your picture and applied to the original pixels. A 1200×900 photo is saved 1200×900, whether as PNG, as JPG, or inside the ZIP.
Can I save as JPG?
Yes, at quality 0.92 — but only once the subject has something behind it. JPG has no alpha channel, so saving a transparent cut-out as JPG would fill the hole with black without asking. Pick a colour or a photo for the background and the JPG option unblocks itself.
What file types and sizes does it take?
PNG, JPG, WebP, GIF (its first frame) and BMP, up to 40 MB per file — the same 40 MB ceiling applies to a background photo, since that is decoded the same way. A file that is over the limit, or is not a picture at all, is named on screen instead of being silently ignored.