Create the cutout first, then make the edge yours
Choose a JPG, PNG, or WebP image and select Remove background. The first run loads about 6.6 MB of model data from NexaCurrent and analyzes the picture on your device. The model estimates which pixels belong to the main subject and turns that estimate into an editable mask. Use After to inspect the cutout and Before to return to the source. Transparent creates a PNG that can sit over another design. Solid color places the cutout over the color you choose, which is useful for profile cards, simple listings, slides, thumbnails, and social graphics.
The automatic result is only a starting point. Select Restore subject and brush over parts removed by mistake, such as hair, fingers, clothing, glasses, or a held object. Select Erase background to remove leftover wall, floor, sky, or edge halos. Use a smaller brush around fine edges and a larger one for broad corrections. Reset brush changes returns to the automatic mask without reopening the source or loading the model again.
Choose photos the model can separate honestly
This first version is strongest on portraits and people with a reasonably distinct background. Soft, even light and clear separation between subject and background usually produce a better starting mask. Hair, fur, transparent fabric, smoke, motion blur, shadows, reflective products, multiple overlapping subjects, and similar foreground and background colors are harder. Product photos, plants, vehicles, illustrations, and groups may work, but they should be treated as drafts that can need substantial brush correction rather than guaranteed one-click results.
Inspect the outline at a larger browser zoom. Check hair tips, hands, gaps, holes in objects, semi-transparent edges, and contact shadows. A contrasting solid background can reveal halos hidden by the checkerboard. If the mask contains almost everything or almost nothing, choose a clearer image, crop closer to the subject, or use a specialist editor. Repeating the same uncertain image will not create missing visual evidence.
Local model loading, cancellation, and safe limits
The model and browser calculation files are hosted by NexaCurrent so the tool does not silently download changing code or weights from another service. Once those files load, the source pixels, mask edits, previews, and export remain on your device. Progress shows the current run, and Cancel stops waiting for the active result; the source image remains available so you can retry or choose another file. A repeated run replaces the editable mask only after a new result finishes, which helps prevent a cancelled or older task from overwriting your current work.
Images are limited to 20 MB and 12 megapixels because decoded pixels, model input, masks, and export canvases can occupy much more memory than the compressed file. Only one image is processed at a time. The output is PNG because transparency needs an alpha channel; a solid background is also flattened into that PNG. The tool does not restore hidden background pixels, create a layered design file, guarantee a commercially perfect clipping path, or determine whether you have permission to use the source image. Keep the original, inspect the downloaded PNG in its destination, and follow the applicable consent, copyright, brand, and disclosure requirements.