Background Remover

What is a background remover?

A background remover separates the main subject of a photo — a person, a product, a pet — from everything behind it, and makes the background transparent. This free background remover does the job with an AI segmentation model that runs entirely in your browser. You get a PNG with a real alpha (transparency) channel that you can drop onto any design, or you can replace the background with a solid color or another photo right here and download the result as PNG or JPG.

Nothing is uploaded: your browser downloads only the AI model (once), and the photo is processed on your own phone or computer — suitable for ID photos, confidential product shots or family pictures.

How to use the background remover

1. Add a photo — click the upload box, drag a file onto it, or copy an image and press Ctrl+V (Cmd+V on Mac). JPG, PNG and WebP work in every modern browser.
2. Pick the AI model — Best quality (ISNet) gives the cleanest, most detailed edges; Fast / small (U2-Netp) is a tiny download that finishes in seconds, at the cost of fine detail. Either one is downloaded only the first time.
3. Click Remove Background — the progress bar shows the model download in megabytes, then the processing step, which runs in a background thread so the page stays responsive.
4. Check the result — drag the before/after slider across edges such as hair, fingers and shadows.
5. Choose a new background (optional) — keep it transparent, pick a solid color (white for marketplaces, blue or white for ID-style photos), or upload your own background image, which is scaled to cover the frame.
6. Download — PNG keeps transparency; JPG gives a smaller file with a solid background.

How the AI background removal works

The tool runs an AI image segmentation model with ONNX Runtime Web (by Microsoft, MIT licence) inside a background worker in your browser. It uses your graphics card through WebGPU when the browser and model support it, and otherwise WebAssembly on your processor. Two openly licensed models are available:

• Best quality — ISNet general-use (“Highly Accurate Dichotomous Image Segmentation”, Xuebin Qin et al., DIS project, Apache-2.0 licence), in a half-precision ONNX build packaged on Hugging Face. It looks at the photo at 1024 × 1024 px.
• Fast / small — U2-Netp, the lightweight version of U²-Net (Xuebin Qin et al., Apache-2.0 licence), ONNX build on Hugging Face. It looks at the photo at 320 × 320 px, which is why it is much faster but less precise.

The steps are the same for both models:

1. Your photo is decoded and, if its longest side is above 2048 px (or 1024 px if you chose the low-memory option), scaled down so phones do not run out of memory.
2. A copy is resized to the model's input size (1024 × 1024 or 320 × 320) and the network predicts, for every pixel, how likely it is to belong to the foreground (0 = background, 255 = subject).
3. That prediction — the alpha mask — is scaled back to your photo's size and written into the alpha channel. Soft values give smooth, anti-aliased edges instead of jagged cut-outs.
4. For a new background, the tool paints the color or image first and draws the cut-out subject on top.

Worked example: we tested a 668 × 396 px product photo of a stethoscope lying on a white desk. With the Best quality model the tool produced a 668 × 396 PNG in which 76.2% of the pixels became fully transparent while the metal chest piece and tubing stayed opaque; it took about 24 seconds on a desktop processor. The Fast model finished in about 3 seconds and made 83.9% transparent, because it also dropped most of the thin tubing. The “Made transparent” figure is this share of pixels with almost zero opacity.

Download size: the Best quality model is 88.1 MB and the Fast model 4.6 MB. The ONNX Runtime engine adds about 2.5 MB of compressed transfer for the processor version (about 4 MB when your graphics card is used). So the first run downloads roughly 91 MB with Best quality or roughly 7 MB with Fast. The model is then kept in your browser's cache storage, so later runs skip the download (clearing site data removes it).

Use cases

• E-commerce product photos — Amazon, Flipkart and Meesho listings usually ask for a plain white main image. Remove the background, choose Solid color with #ffffff and download JPG.
• Profile pictures and passport-style photos — put yourself on a clean white or light-blue background, then size it with our Profile Picture Maker.
• Stickers, logos and thumbnails — transparent PNG cut-outs for YouTube thumbnails, presentations and WhatsApp stickers.

Tips for clean results

• Use a photo with one clear, in-focus subject that contrasts with the background.
• Even lighting helps. Strong shadows on the floor are sometimes kept as part of the subject.
• Crop very wide shots first with our Image Crop tool so the subject fills more of the frame — the AI sees the picture at about 1024 px, so small subjects lose detail.
• On older phones choose 1024 px as the maximum working size; it uses roughly a quarter of the memory.

Limitations & notes

AI cut-outs are good, not perfect. Fine hair, fur, and fly-away strands are often softened or trimmed. Glass, water, smoke and other see-through objects cannot be separated properly because the background shows through them. Low-contrast photos are the hardest case: in our tests the Best quality model removed part of a white doctor's coat against a pale, blurred hospital background, while the Fast model kept the coat but lost thin details such as stethoscope tubing. If one model gets it wrong, try the other. Busy group photos may lose people at the edges. Shadows are not recreated on a new background.

The tool needs an internet connection only for the first download of the engine and model. It needs a current browser with WebAssembly (Chrome, Edge, Firefox, Safari, Samsung Internet). Processing runs in a background worker so the page stays usable. In our tests on a desktop PC using the processor, Best quality took about 24–25 seconds per photo regardless of its size (the AI always works at 1024 × 1024) and Fast about 3 seconds; with WebGPU the speed depends on the graphics card, and budget phones can take longer. HEIC photos from iPhones must be converted to JPG first. Output is limited to the working size (2048 px on the longest side by default).

Frequently Asked Questions

Is this background remover really free and private?

Yes. There is no sign-up, watermark or limit. Your photo is processed by AI inside your browser and is never uploaded to MavexTech or anyone else. Only the AI model (from Hugging Face) and the ONNX Runtime engine (from jsDelivr) are downloaded the first time you use the tool.

Why does the first run take longer?

The first time, your browser downloads the AI model and engine: about 91 MB for Best quality or about 7 MB for Fast. After that the model is loaded from your browser's cache, so the tool starts much sooner.

Does the downloaded PNG really have a transparent background?

Yes. The PNG includes an alpha channel. Background pixels have zero or near-zero opacity and edge pixels are partly transparent for smooth outlines. JPG cannot store transparency, so JPG downloads get a white background unless you choose a color or image.

Can I replace the background with a color or another picture?

Yes. After the background is removed, set New background to Solid color and pick any color, or choose My own image and upload a background photo. The background image is scaled and centred to cover the whole frame.

Why are hair or glass edges not perfect?

The AI decides for each pixel whether it belongs to the subject. Thin hair, fur, transparent glass and subjects with colours similar to the background are hard to decide, so edges there can be soft, cut off or partly see-through. A photo with good contrast gives the cleanest result.

Will it work on my phone?

It works in current mobile browsers, but processing is slower than on a computer and needs free memory. Photos larger than 2048 px are scaled down automatically; if processing fails, choose the 1024 px option and close other tabs.

Can I use the result commercially?

The images you create are yours to use. The models (ISNet and U2-Netp, Apache-2.0) and ONNX Runtime Web (MIT) are under permissive open-source licences. You are still responsible for having the rights to the original photo.

Related tools

Profile Picture Maker · Image Crop · Image Resizer · Image Compressor · HEIC to JPG · Image Format Converter