@imgly/background-removal (In-Browser) vs a Background Removal API: When to Use Which

IMG.LY publishes an npm package that runs a segmentation model in the browser through ONNX Runtime Web (WebAssembly, with GPU acceleration where the browser supports it). You pass an image (URL, Blob or ImageData) and get back a Blob with the background removed; there is also a Node variant for server-side use.

import { removeBackground } from "@imgly/background-removal";

const blob = await removeBackground(file, {
  // model: "small" | "medium" (larger = better edges, bigger download)
  progress: (key, current, total) => console.log(`${key}: ${current}/${total}`),
});
imgEl.src = URL.createObjectURL(blob);

The first call downloads the model and runtime assets — tens of megabytes, depending on the model you choose — and caches them; subsequent calls are local. Check the package README for the current model options, sizes and the licence terms (the package has been published under an AGPL-style licence with a commercial licence available from IMG.LY; verify what applies to your use before shipping).

Download: a model of tens of megabytes on first use is fine on a desktop with fibre and painful on a phone on 4G — many products gate the feature behind a "this may take a moment" screen, which is exactly the friction the API avoids.

Device time: on a recent laptop a photo takes a few seconds; on a mid-range Android phone it can take much longer, and memory pressure can crash the tab on very large images. You are shipping a GPU workload to hardware you do not control.

Quality: the browser models are smaller than server models by necessity. On clean subjects the result is good; hair, glass and low-contrast edges are where a server model has the headroom. Test your own images — the comparison is free on both sides.

Licensing: this is the one that surprises teams. Read the package licence and decide whether your product's licence is compatible or whether you need IMG.LY's commercial terms. An API has no such question; you pay per image and the code that calls it is yours.

Choose in-browser when the image genuinely must not leave the device (medical, legal, children's photos, regulated environments), when users are offline, when you want zero marginal cost at very high volumes of small images and can accept variable quality, or when the feature is a nice-to-have in a desktop-first app. It is also a fine free preview: run the in-browser model for the instant on-screen result, then send the image to an API for the final export.

Choose the API when quality on hair and product edges is the product, when your users are on phones and mobile data, when you need consistent timing (a few seconds regardless of device), when the workload is server-side anyway (catalogues, cron jobs, bots), or when you would rather not ship a large model and its licence. bgclear's endpoint is one call, returns PNG/WebP/JPG at full resolution, and preview-size results (≤800 px) are free without limit; full-resolution images are one credit — $9 for 100, $39 for 500, $129 for 2,000, no subscription (pricing). Never call it from the browser directly, though: the key would ship in your bundle. Proxy through a route on your server — the Next.js and Node guides show a 30-line one.

Instant preview in the browser, final export through the API. The user drops an image, sees a cutout within seconds with no upload (imgly), and clicks "Export HD" — which posts the original to your server route, which calls the API at size=full and returns the high-quality PNG. Previews are free on both sides, so the only cost is the export the user actually asked for.

// preview (client): instant, no upload
const previewBlob = await removeBackground(file, { model: "small" });

// export (client → your server → API)
const form = new FormData();
form.append("image", file);
const res = await fetch("/api/cutout?size=full", { method: "POST", body: form });
const hdPng = await res.blob();

The server route forwards image as image_file with your Bearer key and an Idempotency-Key so a retried export is never charged twice. If you are consolidating after remove.bg's shutdown on 1 December 2026, the API side uses remove.bg's field names — see the migration page. For a self-hosted server model instead of a browser one, the rembg comparison covers that trade-off.