Image Background Remover

• Fully client-side. The segmentation model runs in your browser using WebAssembly via ONNX Runtime Web. Your image is never uploaded, stored, or logged anywhere.
• First run downloads the model. On first use, a segmentation model (~40 MB) is fetched from a CDN and cached by your browser. Subsequent runs load the model from cache in a fraction of a second. This is the trade-off for running AI without a server.
• Processing time. On a modern laptop, inference takes around 2–5 seconds. On a phone, expect 8–15 seconds. This is normal for on-device AI.
• Not perfect on every image. The model does well on portraits, products and clear subjects. Hair, fur, glass, semi-transparent objects and low-contrast photos may need manual touch-up with the edge refinement controls.
• No automatic saving. Your image is held in memory only. Reloading the page clears everything.
• Model licensing. The default ISNet model is provided under a non-commercial license by its authors. For commercial projects, replace the model with one that permits commercial use, or contact the authors.
• Large images. Very large images may be slow or run out of memory on low-end devices. For the best experience, use images under about 4000 × 4000 pixels.

A practical guide to AI background removal

Removing the background from a photo used to require either a careful manual selection in an image editor or a paid cloud service that uploaded your file to a server. In the last few years, that changed. Segmentation models small enough to run in a web browser now produce results that were the exclusive domain of desktop software just a few years ago. This tool uses one of those models, running entirely on your device.

How on-device segmentation works

A segmentation model takes an image as input and outputs a mask — a per-pixel map indicating how likely each pixel is to belong to the foreground. The model used here is a variant of ISNet, an image segmentation network that has been exported to the ONNX format and quantised to reduce its size. The browser runs it through ONNX Runtime Web, a WebAssembly build of the same runtime used by many production machine learning systems.

The first time you use the tool, the model file (about 40 MB) is downloaded and cached. Subsequent runs load it from cache, so they start almost instantly. There is no server round-trip on any future use.

What the model does well, and what it does not

On-device models have improved dramatically, but they are not perfect. They tend to work best on:

• Portraits and people. The model has been trained on many examples of human subjects and handles hair reasonably well.
• Products with clear edges. Objects like bottles, shoes, and packaged goods with distinct boundaries are usually cut out cleanly.
• High-contrast backgrounds. When the subject and background are visually distinct, the mask is more reliable.

They struggle with:

• Hair and fur. Fine strands are difficult because they occupy sub-pixel widths. The result is often a slightly jagged or overly-soft edge.
• Semi-transparent objects. Glass, smoke, and water have no clear foreground/background boundary, so the model has to guess.
• Low-contrast photos. If the subject and background are similar in colour or brightness, the mask can bleed.
• Busy, textured backgrounds. Dense patterns behind the subject can confuse the model.

Improving the result with edge refinement

The raw output of the model is a starting point, not a finished cutout. This tool exposes two sliders that let you tune the result. Edge softness blurs the alpha channel slightly, which hides jagged pixel edges at the cost of some sharpness. Mask threshold shifts the decision boundary: a lower value keeps more of the subject and may include some background; a higher value removes more background and may eat into the subject. The right values depend on the image, so experiment until the edge looks clean.

Compositing and output

The final step is compositing. The mask becomes the alpha channel of the output. Transparent areas stay transparent, and the subject is preserved. You can then choose to keep the transparency or fill the background with a solid colour for use in contexts that do not support alpha, like JPEGs destined for a specific layout. PNG is the correct output format for transparency because it is the only widely supported image format that preserves an alpha channel.

About this tool

This background remover is part of the TempScript free online tools collection. It runs entirely in your browser, requires no account, and never sends your images anywhere. If you work with images regularly, you may also find our image compressor and images to PDF converter useful.