The result is an alpha mask
Automatic removal does not literally erase a color. A segmentation model estimates how visible each pixel should remain. White areas of the mask keep the subject, black areas remove the background, and gray values create semi-transparent edges. Most visible failures come from the mask rather than the original RGB pixels.
Why holes inside an object stay filled
Models often recognize the outer silhouette more confidently than openings inside it. A circular product, chair frame, handle, or mechanical part may be treated as one solid object, especially when the inner background is dark and resembles the product. Inspect every enclosed opening at 200% zoom. If an opening is filled, remove it with a manual erase or cutout tool instead of tightening the entire mask and damaging the outer edge.
Dark subject on dark background
Low contrast makes the boundary ambiguous. Increase exposure slightly before removal, but avoid crushing blacks or clipping highlights. If you can reshoot, use a plain background whose brightness and hue differ from the subject. Edge lighting is often more useful than adding general brightness because it creates separation exactly where the mask needs evidence.
Hair, fur, smoke, and transparent materials
Fine strands need soft partial transparency. A hard threshold produces a cardboard edge; an overly soft mask creates a halo. Transparent glass and smoke are harder because the subject contains background color by definition. For marketplace thumbnails, a clean simplified edge may be acceptable. For compositing, retain the original and use a dedicated editor with edge refinement and color decontamination.
How to judge the export
- View the result over white, black, and a saturated color—not only a checkerboard.
- Zoom into corners, hair, internal openings, and contact shadows.
- Confirm the exported file is PNG or another format that supports alpha; JPEG cannot preserve transparency.
- Open the downloaded file in the destination app instead of trusting only the browser preview.
- Keep the original source so the mask can be rebuilt.
Speed versus accuracy
A smaller on-device model starts faster and protects file privacy, but may lose fine structure. A larger remote model can be more accurate after upload, but introduces network time and a different privacy tradeoff. “Instant” and “best quality” are not always the same setting. Choose based on the image's value, edge complexity, and confidentiality.
Use the right tool
Try the background remover for a quick first pass, then use background composer to test the result against a real destination color. A mask that looks acceptable on transparency may reveal a fringe as soon as a contrasting background is added.