Free AI Background Remover That Runs in Your Browser

Instructions

Open a photo

Drop it on the page, click to choose it, or paste it with Ctrl+V. PNG, JPEG and WebP always work; AVIF and HEIC work when your browser can open them. The orientation saved by the camera is applied. Very large photos are offered a reduced working copy so the tab does not run out of memory; your file is never changed.

Choose the model

Image type and Quality pick the model, and the presets set both together with matching edge settings. Each model is downloaded from this website the first time it is used, checked against its published SHA-256, and kept by your browser for the next visit.

• Fast · People — MODNet (7 MB): portrait matting with soft hair, in about a second. It is for people only and can miss parts of unusual poses. With Balanced or High it serves as the guide for Keep only the person.
• Fast · Objects — U²-Netp (5 MB): for slow devices; rougher edges.
• Balanced — ISNet (179 MB): clean cut-outs of people, products, animals and objects at 1024 px. The default.
• High — BiRefNet-lite (224 MB): the finest edges. It needs a graphics card with WebGPU, because it does not fit in the memory the CPU runtime is allowed; without one, ISNet is used and the page says so.

Processor chooses the GPU (WebGPU) or the CPU. When the GPU fails or runs out of memory, the same model runs on the CPU instead, and the page tells you. On the CPU the larger models take from a few seconds to about half a minute, depending on the computer.

Keep only the person

On photos of people a general model sometimes keeps a faint haze of the background, or an object next to the person, and can miss dark clothes or hair. Keep only the person runs a second, portrait model (MODNet, 6.6 MB) as a guide: Standard keeps the parts of the cut-out the person touches, removes what is apart from them, and adds back hair and limbs the main model missed; Strict also cuts what lies beyond the person even when it touches them, such as a car or a chair behind them — check arms and held objects, and paint back anything it took. It is on by default for the People type and the Portrait preset. Tested on photos of people placed on known backgrounds, Standard removed about three quarters of the background haze and cut the part of the person that was lost from 20 % to 8 %.

Adjust the edge

Every slider works on the model's mask without running the model again, and is measured in pixels of the original photo, so the export matches the preview.

• Threshold keeps more (lower) or less (higher) of the faint areas. 50 is the model's own mask. Transition width (in Advanced) makes the step from removed to kept softer or harder.
• Feather softens the edge. Expand / contract trims a halo left around the subject or gives back an edge cut too tight.
• Residue cleanup fades out faint leftovers of the background that lie away from the solid subject; the higher it is, the closer to the subject it reaches, so watch loose hair.
• Halo removal takes the old background's colour out of semi-transparent edge pixels — hair, fur, motion blur — so no green or white fringe shows on a new background.
• Preserve fine transparency keeps soft alpha for hair and fur; off gives a hard, clean edge for rigid objects. Remove small islands and Keep only the largest piece clean up loose specks. Invert keeps the background instead.

Fix details with the brush

Restore (B) paints back what was removed; Erase (E) removes what was kept. Size, hardness and opacity are in the Brush panel; [ and ] change the size. Strokes are kept on top of every slider and of a new model run. Undo with Ctrl+Z, redo with Ctrl+Shift+Z. Zoom with the mouse wheel or a pinch, move with the hand tool or by holding Space; 0 fits the picture and 1 shows it at 100 %.

Background and export

Keep the background transparent, or put a colour or another picture behind the subject. The preview background (checkerboard, white, black or any colour) only helps you see the edge and is never part of the file. PNG keeps full transparency without loss, WebP keeps transparency in a smaller file, and JPEG has no transparency, so a transparent result is put on white. Export at the original size, at 75 % or 50 %, or at a width you type. Mask downloads the mask itself in greyscale.

What no model can do

The mask is predicted at the model's resolution (512 to 1024 px) and scaled to the photo, so the finest strands of hair in a 24-megapixel photo follow the model's detail, not the camera's. Glass, smoke, veils and other see-through things are hard for every model, a shadow on the floor may be kept or removed, and when there are several subjects the model keeps what it finds most salient — use the brush to choose. Nothing hidden behind the subject can be recovered.

Privacy

The page sends nothing about your photo anywhere: not the pixels, the mask, the file name or its metadata. The only downloads are this website's own model and runtime files, all ordinary GET requests you can see in the browser's network panel. The models stay in your browser's storage until you delete them with the button above; photos are never stored.

Credits and licences

• MODNet — Zhanghan Ke et al., Apache-2.0; ONNX export by Xenova (Transformers.js).
• U²-Net / U²-Netp — Xuebin Qin et al., Apache-2.0.
• IS-Net (DIS) — Xuebin Qin et al., Apache-2.0.
• BiRefNet — Peng Zheng et al., MIT.
• ONNX exports of U²-Netp, IS-Net and BiRefNet-lite from rembg (MIT). Runtime: ONNX Runtime Web (MIT).
• Halo removal follows "Approximate Fast Foreground Colour Estimation" (Forte & Pitié, ICIP 2021).