Background Remover

This tool handles one picture at a time: an AI model running inside your browser separates the subject from the background and gives you a PNG with a transparent background. It suits product shots, portraits, pets and single objects that you plan to drop into another design. It does exactly one thing, remove the background. There is no batch mode, no option to swap in a solid color or a new background, no manual brush and no edge-feathering slider, so the quality of the result depends almost entirely on the model, and if it is wrong your only recourse is to try the other mode.

How to use it

1. Before choosing a picture, pick a "Processing mode": "Fast (recommended)" or "High precision". You can also tick or untick "Limit output resolution (longest side under 2000px)". These options disappear once an image is selected; click "Choose again" to change them.
2. Click to select an image, or drag one onto the dashed box.
3. Check the "Original" preview on the left, then click "Start cutout".
4. Wait while the page shows a progress message and an "Elapsed" seconds counter. First use takes noticeably longer because the model has to be downloaded.
5. When done, "Cutout result (high-res)" appears on the right over a checkerboard so you can see the transparency. Click "Download transparent high-res PNG" to save cutout.png.

What the two modes and the output setting actually do

The modes map to two different models. "Fast" uses the smaller "small" model and shrinks the image to a longest side of 1024px before feeding it in; "High precision" uses the "medium" model at 1600px. So high precision uses a bigger model and also looks at more pixels, at the cost of a larger download and slower processing, and it tends to help most on hair and fuzzy edges. The AI computes the transparency mask only on that reduced copy, then the mask is scaled up and applied to the original pixels, so the colors and detail in your result come from the original rather than a blurry thumbnail. The mask itself still comes from a lower-resolution pass, though, so edges can look soft if you zoom in far. With "Limit output resolution" ticked, the result has a longest side of at most 2000px (smaller images keep their size); unticked, the output matches the original dimensions, which makes the compositing step slower and the PNG larger on big photos.

Supported input and known limitations

• Input: any image/* file your browser can open, one at a time. The code sets no file-size cap, but very large images are limited by browser memory.
• Output is PNG with an alpha channel only, always saved as cutout.png.
• First use downloads the model and runtime from a third-party static host (staticimgly.com). The size is in the tens of MB range (larger for high precision), and the time depends on your connection. We found no caching code of our own, so whether it is reused later is down to the browser's normal HTTP cache; clearing site data means downloading it again.
• Inference runs on the CPU via WebAssembly (GPU is not enabled), so on an old device one image can take a long time.
• When the subject and background have similar colors, with translucent objects (glass, gauze), fine strands of hair or intricate cut-outs, the model can leave parts of the background in or cut parts of the subject out. High precision does not guarantee a clean result, and there is no manual tool to fix it.
• If the picture contains several subjects, the model decides what counts as foreground; you cannot choose which one to keep.

Is it stuck? The spinner has been going for a while

Watch the "Elapsed" counter: if it keeps ticking the page is alive. First use must download the model, which can take long on a slow connection. After a minute the page suggests the model download may be blocked and recommends refreshing and retrying. On a network that cannot reach staticimgly.com the tool will not work.

What does "Processing failed" mean?

The message ends with the underlying reason. The most common cause is a failed model download, so check your connection and retry. If the image itself is corrupt or unreadable by your browser, try a different file.

There is leftover background around hair and edges

There is no manual repair. Retry in "High precision", or open the downloaded PNG in your own image editor for touch-ups. Photos with strong contrast between subject and background and even lighting give cleaner results.

Why is the PNG much bigger than my original?

PNG is lossless and carries an alpha channel, so photographic content is usually much larger than the same picture as a JPG. If it is headed for a web page, you can run it through an image compressor afterwards, but remember that converting to JPG throws away the transparency.

Is my image uploaded?

Your image is not uploaded. The code sends no image anywhere; the network is used only to download the model files and runtime into your browser, and the actual processing happens locally.