Image Background Remover — Remove Background Online

The alpha channel: what "removing the background" really changes

A colour photo stores three channels — Red, Green, and Blue — at 8 bits each, giving 256 levels per channel and 16,777,216 possible colours (256 × 256 × 256). Removing a background adds a fourth channel, Alpha, producing the RGBA model. This opacity map is also 8-bit: 256 levels from 0 to 255, where 0 is fully transparent, 255 is fully opaque, and the 254 steps in between draw the soft edge of hair or a blurred sleeve. The neural network deletes no colour data; it computes one new Alpha value per pixel and leaves the tones untouched beneath. That is why the download is a Portable Network Graphics file: the PNG-24 format carries a true 8-bit Alpha channel, while a Joint Photographic Experts Group image stores only 3 channels at 8 bits and cannot record transparency at all.

The grey-and-white checkerboard behind the result is a rendering convention, not image data — Adobe Photoshop, Affinity Photo, and Google Chrome all draw the same 20-pixel checker to signal "0 Alpha here." The genuinely hard pixels are the 1–2 along the boundary. A bottle shot against a plain #FFFFFF wall gives a crisp edge; 3,000 fine hair strands against a cluttered background force the model to estimate a fractional Alpha for each edge pixel, and a wrong estimate of even 10 to 20 levels out of 255 shows up as a faint halo once the cut-out lands on a new colour.

ISNet, ONNX Runtime, and WebAssembly: the actual stack

This tool runs the ISNet segmentation model through the open-source @imgly/background-removal library, version 1.7.0. ISNet — short for Dichotomous Image Segmentation Network, introduced in the 2022 Dichotomous Image Segmentation paper — is a newer architecture than the U2Net model many older tutorials cite. It ships as an Open Neural Network Exchange file and is executed by ONNX Runtime Web, compiled from C++ to WebAssembly so it runs at near-native speed inside the browser tab rather than on a server. The default variant is isnet_fp16, a 16-bit half-precision model of about 80MB; a quantised 8-bit variant near 40MB trades a little edge accuracy for a smaller download.

ISNet does not colour-key or edge-detect. Trained on the DIS5K dataset of roughly 5,470 high-resolution labelled images, it predicts, for every one of the millions of pixels in your photo, a probability from 0 to 1 that the pixel belongs to the foreground. Those probabilities become the 256-level alpha mask. Because it reasons over the whole frame, it separates a subject wearing the same navy as the wall behind them — a case simple #RRGGBB thresholding cannot solve. It remains a prediction, so on a cluttered 12-megapixel scene it can pick the wrong subject or trim 2–3 pixels it should have kept; the cleanest input is one clear subject with good tonal separation.

One click to a finished background, not just a cut-out

A bare transparent PNG is only half of most jobs — an Amazon listing needs the product on #FFFFFF white, a LinkedIn profile photo often wants a flat brand colour, a YouTube thumbnail wants #000000 black. Rather than sending you to a second editor, the Background picker composites the cut-out for you. Pick Transparent, White, Black, or Custom, and the result redraws in under 16ms; Custom opens a full 24-bit colour picker across all 16,777,216 Red Green Blue values.

The compositing runs on an HTML5 canvas at the cut-out's exact pixel dimensions — a 2000 × 1500 input is filled with your chosen colour and the subject is painted over it at the same 2000 × 1500, so no resolution is lost. Transparent choice skips the canvas entirely and hands back the original PNG with its alpha channel intact. Download PNG then saves whichever version is on screen, named background-removed-transparent.png or background-removed-on-white.png so the two are easy to tell apart in a folder of 50 exports.

Your image stays local; only the model is fetched

The important line is between your image and the model. Your image is decoded and run through ISNet entirely on your own device, inside the browser tab — 0 copies of the photo are uploaded to any server. That privacy benefit is real: you can cut out a private photo without it landing on a third party's disk, unlike Remove Bg or Adobe Express, which upload every file to a remote server first.

The model is the one thing that crosses the network. On the first run, roughly 40–80MB of Open Neural Network Exchange weights and WebAssembly are fetched from the IMG.LY CDN at staticimgly.com— a normal asset request, the same kind a page makes for a Google Font or a script, and that CDN sees the request but never your image. The browser then caches those files under Cache Storage, so the 2nd, 3rd, and every later removal need 0 network requests and work fully offline. Claiming "100% offline, 0 requests" would be false on run 1; the accurate statement is that your picture stays private while an 80MB reusable model is downloaded once and reused for the next 100 cut-outs.

Cleaner results, memory limits, and when to use Photoshop instead

Three inputs raise the hit rate. Start from the least-compressed file, because Joint Photographic Experts Group artefacts at quality 60–70 blur the exact 1-pixel edge ISNet needs. Favour tonal contrast between subject and backdrop. And keep the subject sharp — motion blur and a shallow f/1.8 aperture soften that boundary. Placing the cut-out onto a saturated colour? Preview it there first: a 15-level halo invisible on the checkerboard can surface against a bright #22C55E green or a #1D4ED8 blue.

Mind the memory ceiling, too. Each frame is held uncompressed as RGBA at 4 bytes per pixel — a 24-megapixel photo occupies about 96MB before inference even starts — so files past 10MB, or resolutions above 4000 × 4000, can crash the tab on a phone or a 4GB laptop; downscaling to 2000px on the long edge fixes almost every failure. For a catalogue of 500 products, a portrait with fine flyaway hair, or an A3 print at 300 DPI, treat this as a 5-second first pass and finish the hardest 5% of edges with a manual layer mask in Adobe Photoshop, Affinity Photo, or GIMP. For a LinkedIn headshot, an Instagram post, a Slack avatar, a Google Slides graphic, or a quick mock-up, the automatic cut-out is usually the finished article — and 100 of them cost $0 and upload nowhere.

Quick reference

• Input formats: JPG, JPEG, PNG, WebP, up to 50MB per file.
• Output format: PNG-24 carrying 8-bit transparency (256 opacity levels, 0–255).
• Model: ISNet, variant isnet_fp16, 16-bit half-precision, ~80MB; quantised 8-bit variant ~40MB.
• Library: @imgly/background-removal 1.7.0, executed by ONNX Runtime Web on WebAssembly.
• Model source: IMG.LY CDN (staticimgly.com), fetched once, then cached for 0-request offline reuse.
• Backgrounds: Transparent, White (#FFFFFF), Black (#000000), or any of 16,777,216 custom Red-Green-Blue colours.
• Compositing: HTML5 canvas at native resolution — a 2000 × 1500 input exports 2000 × 1500.
• Practical limits: best under 10MB and 4000 × 4000; downscale to 2000px on the long edge if a large file fails.
• Privacy: your image is processed on-device and never uploaded; only the reusable model crosses the network.
• Cost: $0, unlimited, no watermark, no sign-up.
• Cached timing: roughly 3–30 seconds, scaling with megapixels and CPU.
• Best subjects: 1 person, 1 product, or 1 animal, tonally distinct from the backdrop.