AI Background Remover

An AI background remover finds the subject of a photo, whether a person, a pet or a product, and makes everything behind it transparent. This one runs a segmentation model inside your browser tab. The photo is never uploaded, and the cutout comes back at the resolution you gave it, so a 4000 pixel product shot stays 4000 pixels wide instead of being shrunk to a preview size until you pay. The one exception is a very large photo on a phone or low-memory computer, which is reduced just enough to fit in the browser's memory, and the result says so.

Two open models do the work, and the page picks one for you. In Chrome or Edge on a computer with WebGPU, it runs BiRefNet, a high-resolution segmentation network released under the MIT licence, on your graphics card, twice: once on the whole photo, then again zoomed in on the subject, so the edges get the full resolution of the network. Elsewhere it runs IS-Net general-use (Apache-2.0) on the processor. The model downloads once, about 98 MB for BiRefNet or 46 MB for IS-Net, and is kept in the browser cache, so the second photo and every visit after that start straight away, even offline.

Most free removers stop at the raw model output, and that is where two familiar faults come from. The first is specks: off-white or cream patches of background that the model was unsure about survive as faint islands. After the model runs, this tool keeps only the connected regions that make up the subject, plus a narrow band around them for hair, and drops the rest. The second is a dark or coloured shadow inside hair. A strand that is 40% opaque is also 60% old background, and copying the original colour under the new transparency keeps that 60%. The tool estimates the true colour of every partly transparent pixel (the blur-fusion method from Forte and Pitie, 2021) and replaces it, so hair looks right on white, black or any new background.

After the cutout, you choose what the file should look like. Keep it transparent, or put the subject on a solid colour, on a photo of your own, or on a blurred copy of the original for a portrait effect. Resize by percent (10 to 400%) or to an exact width and height, crop to the subject with padding, and save as PNG, WebP or JPG. The cutout can also be enlarged with AI upscaling, which rebuilds detail at 2 or 4 times the size with Real-ESRGAN. The subject is found once on your original photo, so changing the upscale never redoes the cutout.

It has limits worth knowing before you start. The first run waits for the model download, which takes 10 to 30 seconds on a typical connection. AI upscaling needs WebGPU; 2x is offered for images up to 4 megapixels and 4x up to 1.2 megapixels. Glass, smoke and other see-through subjects come back partly transparent, because they are. A subject that is almost the same colour and texture as the wall behind it can still lose a piece of its edge, and the before and after slider is there so you can check that before downloading.