Remove Background
Choose a photo and press Remove background. A segmentation model (U²-NetP, MODNet or U²-Net) is downloaded from this site once, stored in your browser, and runs locally with WebGPU or WebAssembly. You get a transparent cut-out you can refine with edge controls and a restore/erase brush, place on a colour, blurred or custom background, and download as PNG or WebP.
This tool runs in an isolated frame on tools.freetoolspro.com. Your files and inputs stay on your device, and advertising on this page cannot read them.
The three modes
Fast uses U²-NetP (4.6 MB), a compact salient-object model run at 320×320 and scaled up. Portrait uses MODNet (25.9 MB), a portrait-matting model run at about 512 px on the short side; it is meant for people, not objects. Quality uses the full U²-Net (176 MB) at 320×320; it needs more memory and download time. On 300 real photos of general objects from the DUTS-TE benchmark, Quality matched the labelled masks more closely than Fast (mean IoU 0.78 vs 0.75; a statistically significant difference, though Fast was better on 41% of individual photos) at about 2.7× the processing time. On 1,000 labelled portrait photos (the P3M-500-P and P3M-500-NP benchmarks), Portrait mode was clearly the most accurate for people: mean IoU 0.95–0.97 versus 0.89–0.94 for the other modes, with lower error along hair and fine edges, and the same advantage on plain and busy backgrounds. On non-people photos it performed poorly (mean IoU 0.36), so choose Portrait for people and Fast or Quality for everything else.
What stays on your device
Your photo, its file name and the result never leave the isolated tools.freetoolspro.com frame, whose security policy only allows requests to this site. The only download is the public model file, cached in your browser so later uses work offline. “Delete downloaded models” removes it.
Requirements and limits
Built for current Chrome, Edge, Firefox and Safari (verified so far in Chromium-based browsers). WebGPU speeds it up where available; otherwise WebAssembly runs on the CPU. Images over 16 megapixels are reduced for processing. Low-memory phones may not be able to run Quality mode. Fine hair, glass, smoke and low-contrast edges may need touch-ups.
Measured speed (one desktop, 27 Oct 2026)
On an Intel Core Ultra 7 265K with integrated graphics in Chromium, processing time per image after the model was loaded: Fast ≈ 0.9 s (WebAssembly; the U²-Net graphs use an operation ONNX Runtime’s WebGPU backend cannot run yet), Portrait ≈ 0.3 s (WebGPU) or ≈ 1.1 s (WebAssembly), Quality ≈ 2.4 s (WebAssembly). Phones and older computers will be slower; we have not yet measured them, and quality has not yet been compared on a labelled photo set.
Credits and licences
U²-Net and U²-NetP: Qin et al., “U²-Net: Going Deeper with Nested U-Structure for Salient Object Detection”, Pattern Recognition 2020, Apache-2.0 (source); ONNX files from rembg (MIT). MODNet: Ke et al., AAAI 2022, Apache-2.0 (source); ONNX conversion Xenova/modnet (Apache-2.0). Inference by ONNX Runtime Web (MIT). Full notices: third-party-notices.txt. Benchmarks used DUTS-TE (Wang et al., CVPR 2017) and P3M-500 (Li, Ma, Zhang and Tao, “Privacy-Preserving Portrait Matting”, ACM Multimedia 2021, MIT licence); no dataset images are published on this site.
Common questions
Is it really free and private?
Yes. There is no upload, account or credit system; processing uses your device’s CPU or GPU.
Why is there a download the first time?
The AI model has to be on your device to run there. It is fetched once from this site and stored by your browser.
Which mode should I use?
Portrait for photos of people: it was the most accurate on our portrait benchmark. Fast for products, objects and animals; Quality if object edges need a little more accuracy and your device has memory to spare.
Methodology reviewed 2026-10-27. Processing details and test approach.