Remove background from image — free, full resolution, in your browser

Drop a photo. The cutout is computed on your device, so the image never leaves it.

Cut out an image

Drop a photo here

or paste it with ⌘V / Ctrl+V. JPG, PNG or WebP, up to 12 megapixels, five at a time.

The first cutout downloads the {mb} model ({other} if your browser has no WebGPU). After that it stays in the browser cache and this page works offline.

This page is built for a desktop browser: it downloads 94–183 MB of weights and keeps the whole image in memory, which most phones will refuse. Open it on a computer, or use ConceptCut on the Mac.

Try one of these first

Three CC0 photos, each with a different kind of edge. One click loads it and runs the model — no file of yours needed to see what this thing can do.

How it works

  1. Drop, paste or pick a fileAnything your browser can decode — JPG, PNG, WebP — up to 12 megapixels and 20 MB. The file is read straight from disk into this tab; there is no upload step to wait for.
  2. BiRefNet_lite reads a 512×512 copyIt returns one alpha value per pixel. That mask is scaled back up to your image’s real size and applied to the original pixels, so the colours you download are the ones your camera recorded.
  3. Take the PNG, or put a colour behind itA transparent PNG at the original dimensions, or flatten it onto white, black or any colour and save a JPG. The backdrop and the feather slider are applied here in the page — moving them does not re-run the model.

The model on this page

The cutout comes from BiRefNet_lite, trained by Zheng Peng and colleagues on the DIS5K dataset, and re-exported to a fixed 512×512 ONNX graph by studioludens. That re-export is the reason this page exists at all: every 1024×1024 BiRefNet export runs out of memory in the browser, on WebGPU and WebAssembly alike. It runs through Transformers.js on ONNX Runtime Web.

Model
BiRefNet_lite, 512×512 ONNX export by studioludens
Licence
MIT — commercial use allowed, attribution kept on this page
One-time download
94 MB on WebGPU (fp16) · 183 MB on CPU (fp32)
Measured here
0.7 s per image on WebGPU, 5.5 s on the single-threaded CPU path (M2 Max, 1600×1066)

Read the MIT licence text that ships with this page →

Questions people actually ask

Do my photos get uploaded anywhere?

No. This page has no upload step at all. Your browser downloads the model file once — 94 MB with WebGPU, 183 MB in CPU mode — and from then on every cutout is computed inside the tab. You can check it rather than trust it: open the Network panel, cut out a second photo, and watch that no request leaves for the image.

Is the free version really full resolution?

Yes. The PNG you download has exactly the pixel dimensions of the file you dropped in: a 4000×3000 photo comes back 4000×3000, no watermark, no daily count. remove.bg’s free tier hands you a 0.25 MP preview — roughly 625×400 — and charges for the real file. The limits here are 12 megapixels per image and five images per drop, and both are browser memory limits, not a paywall.

How good are the edges — hair, fur, glass?

Good at outlines, fair at single strands. The model sees a 512×512 version of your photo, so fur and thick hair come back convincing (the corgi example keeps its ear fluff and most whiskers) while individual flyaway strands and the gaps between them get simplified. Semi-transparent things — glass, veils, smoke, motion blur — are treated as solid, because the output is one alpha value per pixel rather than a true matte. The 0–3 px feather slider softens the stair-stepping left by the upscale.

Why does the first run download 94–183 MB?

Because the model has to come to your machine — that is the whole trade. WebGPU uses the 16-bit weights (94 MB); CPU mode needs the 32-bit ones (183 MB) because the 16-bit graph will not build there. It happens once: your browser keeps the file in its cache, so a second visit is ready in one to three seconds and works with the network unplugged. Clearing site data is what makes it download again.

What if my browser has no WebGPU?

It still works. The page falls back to WebAssembly, which is the same model running on one CPU thread: about 5.5 seconds per image on an M2 Max, and longer on an older laptop, instead of about 0.7 seconds on a GPU, after a larger 183 MB download. Safari and Firefox land here today. You can force either route by adding ?device=wasm or ?device=webgpu to the URL.

Can it do batches, or video?

Batches of up to five images per drop, yes — they queue up, you can flip between them and download them one by one or all at once. Video, no: this is a single-frame model and running it thirty times a second in a browser tab is not realistic. For whole folders, for cutting by concept instead of by the one salient object, or for erasing things inside the picture as well as behind it, that is what the Mac app below is for.