Cut out people and objects in your browser — nothing gets uploaded
Drop an image here or click to choose
JPG, PNG, or WebP · processed on your device
| Model | License | Download size | Best at | In this tool |
|---|---|---|---|---|
| MODNet (full precision) | Apache-2.0 | 24.7 MB | Portraits, people | Yes — Balanced mode |
| MODNet (8-bit quantized) | Apache-2.0 | 6.3 MB | Same, slightly softer edges | Yes — Fast mode |
| U2-Net | Apache-2.0 | ~168 MB (fp32) | General objects | No — too heavy for a page load |
| BRIA RMBG-1.4 | Custom, non-commercial | ~44 MB | General purpose | No — license forbids commercial use |
License matters as much as quality here. Several popular background-removal models carry non-commercial terms, which is why they're excluded even though they benchmark well. Everything this tool loads is Apache-2.0, both the runtime (transformers.js) and the weights (MODNet).
Old-school background removal drew a hard boundary around a subject. Modern matting models do something better: for every pixel they predict an alpha value between 0 and 1 — how much of that pixel belongs to the person and how much belongs to the background. Semi-transparent pixels along hair and shoulders stay semi-transparent, which is what makes cutouts look natural when you paste them somewhere new.
Your image is decoded, resized to what the model expects, and pushed through MODNet as a tensor. The model outputs a single-channel matte; transformers.js multiplies it into an alpha channel, reattaches your original colors, and hands back a full RGBA image. From there the page composites it onto transparent, white, or a color you pick, and encodes a PNG.
Drop a photo, pick a quality setting, and wait for the matte. First run downloads the model; after that it's cached and each image takes seconds. Use the output-background selector to flatten onto white for a clean product-card look, or leave it transparent and drop the PNG into a slide deck, an email, or a marketplace listing.
Take a 1920×1080 photo. That's 2,073,600 pixels (2.07 megapixels). Before the model runs, the raw color data is 1920 × 1080 × 3 bytes = 6,220,800 bytes, about 5.93 MiB. The output adds an alpha channel: 1920 × 1080 × 4 = 8,294,400 bytes of raw RGBA, 7.91 MiB — a third more data, because each pixel now carries opacity alongside red, green, and blue. PNG compression typically squeezes that back down below the original JPEG's size for clean cutouts with large empty regions.
On the model side, the trade is the download: full-precision MODNet is 24.7 MB, while the 8-bit quantized build is 6.3 MB — 25.5% of the original size, a 74.5% saving, at the cost of slightly softer edge gradation. If you're processing one photo on a phone, Fast mode is usually the right call; for headshots going on a company page, let Balanced load once and cache.
Yes. The tool downloads the open-source MODNet model (Apache-2.0) straight from a public CDN and runs it on your device with WebAssembly. Your photo is never uploaded to any server — you can watch the network tab and see that nothing leaves the page after the model files arrive. There is no watermark, no signup, and no image limit because there is no server doing the work.
The first removal downloads the model weights — about 25 MB for the full-precision version or 6.3 MB for the quantized one — plus the WebAssembly runtime. Your browser caches both, so the second and later images process in seconds. If you switch between the Balanced and Fast quality settings, each setting's model is downloaded once.
MODNet is a portrait matting model, so photos of people give the cleanest cutouts. Product shots, pets, and objects with clear edges usually work well too. Fine details like individual hair strands, glass, and sheer fabric are where free in-browser models fall short of paid services — expect soft edges there instead of perfect ones.
It doesn't — that checkerboard is only the on-page preview showing you where the image is transparent. The exported PNG keeps true transparency. If a program shows a solid white background instead, that app filled transparent pixels for display; the alpha channel is still in the file.
PNG. JPEG has no alpha channel, so a transparent cutout saved as JPEG gets a solid background baked in. PNG and WebP both support transparency; PNG is the safe universal choice for pasting into slides, docs, and editors.
remove.bg runs a large server-side model, which is why it handles complex hair and edges better. This tool trades some edge quality for privacy and price: the model runs on your machine, so nothing is uploaded, there are no credits, and there is no resolution cap imposed by an upload limit. For profile photos, product shots, and presentation cutouts, the results are usually indistinguishable.