Background Remover
Runs 100% in your browser — but it downloads an AI model first
Cutting out a background needs a machine-learning segmentation model. The first time you press Remove Background, that model is downloaded into your browser ( for ) and then runs locally on WebAssembly or WebGPU. On a slow connection the first run can take a minute. Your browser caches the model afterwards, so every later run starts immediately — even offline.
Your image never leaves your device. There is no upload and no server-side processing: the photo is decoded, segmented and rendered entirely on your own machine. Only the model file itself is fetched from the network.
2. Background
3. Manual Touch-Up
Preview
Features
- Nothing is uploaded: the segmentation model runs on your own machine through WebAssembly (and WebGPU where available). Your photos are never sent to a server.
- Modern segmentation models: RMBG-1.4 handles general subjects — products, animals, objects and people — while MODNet is a lightweight option tuned for portraits.
- Cached after the first run: the model file is stored by your browser, so repeat runs start instantly and keep working with no connection.
- Manual touch-up brush: erase leftover background or restore parts the model trimmed too aggressively, right on the preview.
- Transparent or solid background: export a transparent PNG, or drop the cut-out onto any solid colour and save it as PNG or JPG.
- Adjustable edge softness: feather the cut-out edge so composites blend naturally.
How to Use
- Drag an image onto the upload box, or click it to browse your files.
- Pick a model. RMBG-1.4 is the best all-round choice; MODNet is smaller and faster if your subject is a person.
- Click Remove Background. The first run downloads the model — watch the progress bar — and subsequent runs skip straight to processing.
- Optionally choose a solid background colour and adjust Edge Softness.
- Use Erase and Restore to paint corrections directly on the preview.
- Download the result as a transparent PNG, or as a flattened JPG.