The five steps of every image
- The browser validates the image
Choose files, drag them in, paste from the clipboard, or start with a sample. BGCut accepts PNG, JPG, WebP, HEIC, and HEIF up to 20MB each, ten images per batch. The server checks the file again — actual decodability, not just the extension — so damaged uploads fail with a clear error instead of a bad result.
- A model estimates the subject mask
RMBG (the default) or BiRefNet analyzes the image and estimates which pixels belong to the foreground. The output is a soft alpha mask, not a rectangular crop, so hair, fabric, and anti-aliased edges keep partially transparent pixels. Processing time, output size, and peak GPU memory are reported for every run.
- Preview backgrounds expose edge problems
A cutout can look fine on the checkerboard while carrying a pale halo from its old background. Switch between transparent, white, and soft-gray previews, and drag the before/after slider. Previews never change the exported file.
- Smart and manual tools correct the mask
Smart erase removes a connected object near a stroke; smart restore does the inverse. Manual brushes give pixel-level control with adjustable size and hardness. Every operation can be undone, and reset always returns to the model output — experimenting is safe.
- Export as a real transparent PNG
The result downloads as a PNG with its alpha channel intact, plus an optional baked shadow for product shots. Nothing is stored as a permanent gallery, and no account is involved.