Background Remover
The first run loads a 84 MB local model; later runs in the same page reuse it
Add files and press start. Everything runs locally in your browser; nothing is uploaded
How to Use
📖 Tool Introduction
Background removal separates the subject from its surroundings and writes out a PNG with a real alpha channel, ready to drop into a poster, a product page or a design file.
This tool runs a U2Net neural network entirely on your machine. The image is scaled to 1024 by 1024 and fed to the model, which returns a greyscale mask of the same size where white marks the subject and black marks the background. That mask becomes the alpha channel of the result.
Because inference happens inside your own browser, no image is uploaded to any server and the tool keeps working offline. The first run downloads a 84 MB model file, and later runs in the same page reuse it.
Every operation happens on your own device. The images you add never leave the machine, are never uploaded to a server, and leave no copy behind anywhere. That matters because it means you can process pictures containing faces, identity documents or commercial artwork without wondering where they ended up. It also means the speed depends entirely on your hardware: larger images and longer batches take proportionally longer, and the tab needs to stay open while the work runs. If you close it partway through, nothing has been lost except the time already spent, because nothing was ever sent anywhere in the first place.
What comes out is meant to be used. Download the files, drop them into whatever you are building, and come back whenever you need another pass. Because there is no account and no upload queue, there is no limit on how many times you run the tool or how many files you put through it, and no waiting behind other people jobs. The one thing worth checking is that the settings you chose actually suit your case: a quality preset that is right for a website hero image may be far too aggressive for a print file, and a crop ratio that flatters a landscape will cut the head off a portrait. Run a single image first, look at the result, then commit the whole batch.
✨ Key Features
- Neural network inference on your device, nothing uploaded
- Three outputs: transparent, solid colour or blurred background
- Adjustable edge handling and alpha threshold for hair and fine detail
- Batch processing with per-image preview and individual download
- PNG or WebP output, both with an alpha channel
- Preview every result, download individually or all at once
🚀 How to Use
- Click or drag to add images; you can select several at once
- Choose how to treat the background and pick a replacement colour
- Tune edge handling and the alpha threshold to suit your image
- Press start; the local model loads automatically on the first run
- Preview each result in the panel and download individually or all at once
⚠️ Things to Note
- Model runs locally: Inference happens entirely inside your browser, so no image is uploaded to any server and the tool keeps working offline
- First load size: The first run loads a 84 MB local model file, so it takes a moment depending on your connection and device; later runs in the same page reuse it
- Limits with hard cases: Transparent objects, subjects close to the background colour and very fine hair may not separate cleanly; adjusting edge handling and the alpha threshold often helps
- What works best: Portraits, products and animals with a clear subject against a contrasting background work best; images where the subject blends into the background are harder
- Choosing an output format: PNG has the widest compatibility while WebP is smaller but unsupported by some older software; both keep a full alpha channel