Remove / Replace Background
100% private — runs on your device, never uploaded. Works offline once loaded.
Cut the subject out with on-device AI, then keep a transparent PNG or place it on a solid colour / new backdrop. No Pro credits or uploads.
How AI background removal works
This tool runs a segmentation model — a neural network trained to classify every pixel in an image as either 'subject' or 'background' — directly inside your browser. The model outputs a mask (essentially a probability map of which pixels belong to the foreground), which is then used to make the background pixels transparent while keeping the subject intact, producing a PNG with an alpha channel.
Unlike older background removal methods that relied on a solid-color backdrop (chroma keying, the 'green screen' technique), this approach works on ordinary photos with any background because the model has learned what people, animals, products and other common subjects generally look like.
What it's typically used for
Common uses include isolating a product photo for an online store listing, cutting a person out of a photo to place on a new background for a profile picture or ID-style photo, preparing a logo or graphic element for a design composite, and creating stickers or overlays from a photo for presentations or social posts.
Where the mask can struggle
Segmentation quality depends on how clearly the subject stands out from its surroundings. Fine, thin detail — individual hair strands, fur, lace, or semi-transparent fabric — is the hardest thing for any background-removal model to cut cleanly, and edges there may look slightly soft or stepped rather than pixel-perfect. Low-contrast boundaries (a subject wearing colors close to the background) and cluttered scenes with multiple overlapping objects also reduce accuracy.
- Sharp, well-lit photos with a clear subject-background contrast segment best
- Fine hair, fur and transparent materials are the hardest edges to get clean
- The output is always a PNG, since transparency requires an alpha channel JPEG doesn't support
- Very large images take longer, since the model processes every pixel
Model download and device performance
The segmentation model is roughly 5 MB and downloads to your browser the first time you use the tool, then stays cached for future use without re-downloading. Because the whole model runs on your device rather than a remote server, processing time depends on your device's hardware — a recent laptop or phone handles typical photos in a few seconds, while older or lower-powered devices may take noticeably longer on large images.
Frequently asked questions
Does the AI model download?
Yes — on first use the model is downloaded to your browser from Whiztools and cached.
Are my images uploaded?
No — background removal runs entirely on your device.
Can I replace the background too?
Yes — choose Replace with colour or Replace with image instead of Transparent PNG.
What file format is the result?
Always PNG — transparency requires an alpha channel, which JPEG doesn't support, so the output preserves the cut-out subject with a transparent background.
Why are the edges around hair or fur not perfectly clean?
Fine, thin detail like individual hair strands is the hardest case for any segmentation model; edges there may look slightly soft rather than razor-sharp.
Can I remove the background and replace it with a solid color instead of transparency?
The tool outputs a transparent PNG; you can then layer that PNG over any color or image you want in an editor or design tool.
Does it work on photos with multiple people or objects?
It generally keeps everything the model identifies as foreground, so multiple subjects can be preserved, but very cluttered or overlapping scenes reduce mask accuracy.
Why does the first run feel slower than later ones?
The AI model (about 5 MB) has to download and initialize the first time you use the tool; it's cached after that, so subsequent images process faster.
Is there a size limit on the photo I can use?
There's no hard cap, but very high-resolution images take longer to process since the model evaluates every pixel, and performance depends on your device's processing power.
Advertisement