Background Remover

How AI background removal works — without the privacy compromise

For years, removing a background from a photo meant either spending 20 minutes with Photoshop's pen tool or uploading your image to a third-party service like remove.bg. Those services process millions of photos daily on their own servers — every portrait, every product shot, every private moment uploaded to someone else's infrastructure. This tool takes a fundamentally different approach: the AI model downloads once to your browser (~20 MB) and runs locally using WebAssembly. The neural network — an IS-Net segmentation model trained to distinguish foreground from background — executes entirely on your device. Your photo never leaves the tab.

As a concrete example: a 4.7 MB JPEG portrait from a smartphone (4032 × 3024 pixels) typically processes in 3–7 seconds on a modern laptop and outputs a 1–3 MB transparent PNG with a clean cutout. The same image on a phone takes 8–15 seconds depending on the device — still faster than uploading and waiting for a server round-trip. The first use downloads the ~20 MB model bundle (one-time, cached by your browser). Subsequent uses are instant because the model is already on your device.

Why no-upload matters for background removal

Photos you process through background removal are often sensitive — product prototypes under NDA, ID documents being prepared for applications, private portraits, medical reference images, or unreleased creative work. Upload-based background removers (remove.bg, Adobe Express, Canva, and many others) require sending your full-resolution image to a remote server for GPU inference. You have no audit trail, no deletion guarantee, no way to verify the image was purged from cache layers, and no control over whether it was logged for model training or quality evaluation. This tool processes everything locally. The model runs in your browser. The only network request is the initial model download from the IMG.LY CDN — after that, the tool works entirely offline. Zero bytes of your image are transmitted to any server.

The IS-Net model: what is it and why it matters

The IMG.LY background removal library uses IS-Net (Image Segmentation Network), a deep learning model designed specifically for salient object detection — figuring out which parts of an image are the "subject" versus the background. The model shipped in this tool is quantized (isnet_quint8), striking a balance between accuracy and download size. Quantization reduces the model weights from 32-bit floats to 8-bit integers, cutting the download to ~20 MB while maintaining >95% of the original accuracy. The model runs on WebAssembly, meaning it compiles to a near-native binary format that executes in the browser's sandbox at speeds approaching native code. No GPU required — it runs efficiently on CPU via SIMD instructions available in modern browsers.

Use cases: from e-commerce to creative workflows

The most common use cases for background removal are e-commerce product photography (cutting out items for white-background listing images), portrait editing for social media or professional headshots, and graphic design where subjects need to be composited onto new backgrounds. Other uses include creating stickers or emoji from photos, preparing ID-style portraits, isolating signatures, and cleaning up document scans. Because the output is always a transparent PNG, you can drop the result directly into any design tool, presentation, or web project without additional editing.