Clean cutouts without sending photos away.

No. The selected image is read and processed by the background-removal model in your browser. Polr Labs does not upload the image to an application server, save a copy, or attach it to an account. The temporary original and result previews disappear when their browser URLs are cleared or the page is closed.

The browser downloads an ONNX background-removal model and the runtime files needed to execute it. Those files describe the machine-learning model; they do not contain or receive your selected photo. Model files are served from the Polr Labs model host and may be stored in the browser cache for later visits.

The first run has to download, initialise, and prepare the model before it can analyse the image. Later runs can be faster when the browser keeps those model files in its cache. Download speed, device memory, processor performance, and image dimensions all affect the wait time.

You can choose a PNG, JPEG/JPG, or WebP image up to 20 MB and 12 megapixels. A 4K image normally fits within the pixel limit. Files with a supported extension can still be rejected if the browser cannot decode the image or if its reported file type does not match the actual content.

PNG supports an alpha channel, which records transparent and partially transparent pixels around the extracted subject. JPEG cannot preserve that transparency. The PNG can be placed over a new colour, photograph, presentation, marketplace listing, or design without restoring the removed background.

It can run in a current mobile browser that supports the required WebAssembly features and has enough available memory. Large images and older devices may process slowly or cause the browser to close the task. If that happens, use a smaller image or a desktop browser with more memory.

The first use requires a network connection to download the model. A repeat visit may work with fewer network requests when the browser has retained the required files, but offline operation is not guaranteed because private browsing, storage cleanup, browser settings, or a model update can remove or invalidate the cache.

Use a sharp image with one clear foreground subject and visible separation between the subject and background. Even lighting and a reasonably different colour or tone behind the subject help the model identify the boundary. Avoid heavy motion blur, extreme compression, tiny subjects, and backgrounds that visually merge with the subject.

These areas do not have a simple solid edge. Fine hair mixes with background pixels, glass reveals what is behind it, and soft shadows can resemble part of the surface. Automated segmentation has to estimate which partially blended pixels belong to the subject, so complex boundaries may need manual refinement in an image editor.

The removed area may be similar in colour, texture, or brightness to the background, or it may be too small for the model to distinguish reliably. Try a higher-quality source, crop closer to the main subject without cutting it off, or use an image with clearer lighting and separation.

Busy objects, reflections, foreground clutter, and background elements touching the subject can be classified as part of the subject. Try a cleaner source image or remove the remaining area manually after downloading. Reprocessing the exact same file is unlikely to produce a different result.

Check the network connection, reload the page, and try again with the tab kept open. Privacy extensions, strict content blockers, restricted networks, or disabled browser storage can prevent model files from loading. A current version of Chrome, Edge, Firefox, or Safari is the best fallback.

Confirm that the file is PNG, JPEG, or WebP, no larger than 20 MB, and no more than 12 megapixels. Then try a smaller copy and close memory-heavy tabs. If the page reports that the file cannot be read, export it again from an image editor to repair unusual or damaged encoding.