AI Background Remover & Mask Studio
Local onlyRemove and replace image backgrounds locally with repairable alpha masks and no image upload.
Images never leave this deviceOnly registered model weights are downloaded; image decoding, inference, masks, edits, and export stay in this browser.
Drop an image into the private mask studio
Open a portrait, product, animal, or other image. Automatic results remain editable instead of being treated as final.
JPEG · PNG · WebP · browser-supported AVIF · up to 100 MB and browser Canvas limits
How to use it
- 01
Open an image locally
Choose, drop, or paste a supported image. The browser checks its real signature and decodes it without uploading the file.
- 02
Generate and refine the mask
Explicitly load a versioned first-party model, run it in a Worker, then repair keep and remove areas or tune the edge without rerunning inference.
- 03
Compose and export
Keep transparency, add a color, gradient, or local image background, then download a real PNG, WebP, JPEG, mask, or project file.
Frequently asked questions
Are my images uploaded?
No. Images, masks, brush edits, previews, and exports stay in this browser. The only optional network download is the registered model file from assets.nobuf.com.
Why is a model downloaded?
Automatic segmentation needs trained weights. They are fetched only after you click the action, checked by byte length and SHA-256, then may be cached locally for later use.
Does it work without WebGPU?
Yes. It tries WebGPU when supported and falls back to ONNX Runtime WebAssembly. Actual support and speed depend on the browser, device, model, and available memory.
Can I fix an imperfect cutout?
Yes. Keep, remove, and erase-correction brushes use source-image coordinates, and threshold, feather, smoothing, and expand or contract controls do not rerun the model.
How does RMBG-1.4 differ from MODNet?
RMBG-1.4 INT8 is a larger general-subject model that can preserve more detail on products, animals, and mixed scenes, while MODNet is smaller and optimized for fast portrait workflows. Results still depend on the source image.
Can I export JPEG with transparency?
No. JPEG has no alpha channel, so a color, gradient, or image background is composed before JPEG export. PNG is the reliable transparent default.
Will it perfectly handle hair or glass?
No model is perfect. Fine hair, transparent objects, motion blur, low contrast, and complex scenes may need brush repair and edge tuning.
Can it process unlimited image sizes?
No. Model inference uses a bounded input, but final masks and exports still depend on browser memory and Canvas limits. Very large files may need resizing first.
Related tools
Keep working with other private browser utilities.