Background Remover Online - Private, No Upload

Local AI image tool

Use Background Remover

The model and image run on this device. The first use downloads about 20 MB from shard.tools; other pages do not load it.

Choose imageJPG, PNG, or WebP; up to 20 MB and 16 megapixels

Best for people and clear foreground subjectsMODNet is a portrait-matting model. Hair and soft edges can work well; glass, fine product gaps, shadows, groups, and busy backgrounds may need manual editing elsewhere.

Explore image tools

Find related browser-local tools for nearby tasks without starting another search.

Browse all image tools

About Background Remover

A background remover separates the foreground subject from the pixels behind it and writes the result with transparency. Most online removers perform that work on a remote server, which means the original portrait, staff photograph, product shot, or family image must be uploaded before the user can even judge the result. This tool takes a different route. It loads a compact MODNet portrait-matting model from the Shard Tools origin and executes it in the browser through WebAssembly. The selected image remains on the device, the model predicts a soft alpha mask locally, and the browser applies that mask to the original-resolution pixels before creating a transparent PNG. The model is lazy-loaded only after Remove background is selected, so visiting the homepage or another tool does not pay the machine-learning download cost. The result keeps soft edge values rather than reducing every pixel to a hard cut, which helps hair and other gradual boundaries look more natural against a new background.

Shape and export choices

Choose a JPG, PNG, or WebP image no larger than 20 MB and 16 megapixels. The browser reads its dimensions and shows a local preview, but no model is downloaded yet. Phone orientation metadata is honoured when the image is decoded, so the output follows the upright preview rather than the camera sensor's raw direction.
Select Remove background. On first use, the page downloads approximately 20 MB of same-origin assets: the quantized MODNet weights and the ONNX WebAssembly runtime. The progress display distinguishes loading from local processing. Browser caching normally avoids downloading those files again on later visits unless the cache is cleared or the model changes.
The source is resized temporarily for model inference while preserving its proportions. MODNet predicts one opacity value per model pixel. That mask is then enlarged smoothly to the source dimensions and applied to the original colour pixels, so the downloadable PNG retains the source width and height rather than saving only the smaller inference image.
Inspect the result over the checkerboard transparency preview. Pay particular attention to hair, shoulders, fingers, glass, gaps between objects, and areas where the subject and background have similar colours. Download the PNG only when those boundaries are suitable for the destination.

Good fits for Background Remover

A private profile portrait

Choose a head-and-shoulders photo with visible separation between hair and background. The local model creates a soft-edged transparent portrait that can be placed on a company colour without sending the employee photograph to an external service.

A simple marketplace product shot

A single object photographed against a plain wall may separate cleanly even though the model is portrait-oriented. Inspect handles, holes, reflective edges, and contact shadows carefully because those details are less reliable than a person's silhouette.

Preparing a circle profile image

Remove the background first, download the transparent PNG, then open Circle Crop Image to frame the face. Each operation remains local, and the second tool preserves the transparent pixels outside the circular crop.

Check the image before download

MODNet is primarily a portrait-matting model, not a universal object segmentation system. A person against a reasonably distinct background is its strongest case. It may also isolate a clear product or pet, but that output is opportunistic rather than guaranteed, and complicated scenes can retain background pieces or remove parts of the subject.
Fine transparent materials, motion blur, flyaway hair, smoke, shadows, reflective surfaces, bicycle spokes, and narrow gaps are difficult for automatic matting. The tool does not include a manual keep/remove brush, so a professional cut-out may still require a dedicated editor after download.
The first run is heavier than ordinary Shard Tools pages because an AI model and WebAssembly runtime must be downloaded and held in memory. Older phones can take noticeably longer or run out of memory on large photographs. The 20 MB and 16-megapixel limits exist to reduce that risk rather than pretending every device has desktop resources.
The output is always PNG because JPEG cannot store transparency. PNG may be larger than the original JPG, especially for detailed photographs. Compressing the result to JPG would restore an opaque background, while converting to WebP can preserve transparency when the destination accepts that format.

Questions about Background Remover

Does the background remover upload my photo?

No. The model files are downloaded from shard.tools, then inference, masking, PNG creation, preview, and download happen in the browser. The site Content Security Policy blocks connections to third-party model hosts.

Why is there a model download on the first use?

Background removal requires a trained image-matting model rather than a simple canvas operation. The quantized model and ONNX browser runtime total about 20 MB. They load only after you start this tool and are normally cached for later runs.

Will it remove every kind of background perfectly?

No. The included MODNet model is designed chiefly for portraits. Clear people and simple foreground subjects work best; glass, groups, intricate products, similar foreground/background colours, and very fine gaps may need manual correction in an editor.

Why is the download a PNG instead of JPG?

PNG supports an alpha channel, which records transparent and partially transparent edge pixels. JPEG has no transparency and would fill the removed area with a solid colour.

Background Remover

Use Background Remover

Explore image tools

Related tools