Account

What new tool do you want?

Image Tools

Remove Background from Image (AI)

Remove the background from photos automatically with AI — people, products, pets, cars. Download a transparent PNG, or swap in a color, blur or your own image. Runs on your device; nothing is uploaded.

How to use

  1. 1Upload a photo — the AI model loads once, then the background is removed automatically.
  2. 2Adjust edge softness or trim halos, and choose a transparent, colored, blurred or custom background.
  3. 3Download a PNG (transparent) or JPG.

Frequently asked questions

Is my photo uploaded to a server?

No. The AI model runs in your browser — on your graphics card via WebGPU when available — so your photos stay on your device.

Why does the first run take a while?

The AI model downloads the first time and is then cached, so later photos are much faster. Automatic mode uses the best-quality model (115 MB) on devices with WebGPU, and a 5 MB model elsewhere.

Which model should I use?

Automatic picks the best one your device can run. Best quality gives the cleanest edges on any subject but needs WebGPU (current Chrome and Edge). Standard works everywhere on any subject. Portraits is tuned for people.

How do I fix a light halo around the subject?

Raise Trim edges to pull the cut-out edge inwards, and lower Edge softness for a crisper outline.

A real segmentation model, not a color trick

This tool is a genuinely different approach from a color-key cutout: it runs an actual machine-learning segmentation model, downloaded once and cached, that outputs a per-pixel transparency mask based on recognizing the subject's shape — not by matching colors. That's why it can separate a person from a cluttered, multi-colored room, something a color-distance tool simply can't do. 'Automatic' picks between three models depending on your device: a 115 MB best-quality model on browsers with WebGPU (current Chrome and Edge), a 26 MB model tuned specifically for people, or a 5 MB general-purpose model everywhere else.

Why edge softness and trim are separate controls

The raw mask from the model is rarely a perfect binary cutout — it often leaves a faint lighter fringe (a 'halo') from the original background bleeding into the edge pixels. 'Trim edges' raises the alpha threshold so borderline pixels are counted as background rather than subject, pulling the cutout boundary inward and removing that halo; 'Edge softness' then applies a blur to the mask edge so the transition isn't a hard jagged line. These solve different problems — a halo needs more trimming, a jagged or overly hard edge needs more softness — and pushing one too far in isolation either leaves a visible fringe or eats into the subject.

Example: swapping a cluttered home-office backdrop

Say you took a product photo of a watch on your kitchen counter, countertop clutter and all, and need it on a plain white background for an online store listing. Upload the photo, let Automatic mode pick its model, and the mask comes back isolating just the watch. Set Background to 'White', nudge Trim edges up slightly if a faint grey halo is visible around the watch band, and download as PNG — the counter and background clutter are gone, replaced with clean white, while the watch's actual reflective edges stay intact.

What 'never uploaded' actually means here

Because the model runs inside a Web Worker in your own browser — on the GPU via WebGPU when supported, otherwise CPU — the image file never leaves your device over the network; there's no server-side step at any point in the pipeline. The trade-off is the first-run download: the model weights (5–115 MB depending on which one loads) have to be fetched once before the very first photo can be processed, after which the browser caches them and later photos process immediately.

Related tools

More: Image Tools