This guide is currently available in English only.
Check the full product outline first
Place the result over both a light and dark solid background. First check whether the product’s silhouette is complete: handles, shoe laces, straps, antennae, chair legs, and negative spaces between parts.
Thin structures disappear when the mask is too aggressive. Background fragments remain when it is too cautious. Both failures are easier to spot against a contrasting color than against a transparency grid.
- Auto-accept only when the expected object count is present.
- Flag an abrupt crop touching the image boundary.
- Compare the non-transparent bounding box with the source subject area.
Review glass and reflective materials separately
Glass, clear plastic, smoke, mesh, and sheer fabric contain information from the old background. Removing that background does not recreate the scene that should appear through the material.
Mirrors, polished metal, and glossy packaging can also reflect the original room. A clean outer edge does not mean the product is ready for a new background.
Route these materials to review by catalog metadata when possible. The reviewer should inspect both the alpha edge and the pixels inside the product—not just the outline.
Decide which objects must remain before processing
A model cannot infer your merchandising policy. A shoe on a display block, jewelry in a box, a three-piece bundle, and a chair with a styled cushion each create a business decision: what should remain?
| Scene | Default action | Why |
|---|---|---|
| Single product, plain support | Remove support | The support is not sold |
| Bundle sold together | Keep every included item | Object count is part of the offer |
| Packaging included | Keep product and packaging | Packaging may be material to purchase |
| Lifestyle prop | Review | The catalog owner must decide |
Review exceptions instead of checking every image
Use product metadata and simple image signals to prioritize review: transparent material tags, multiple detected objects, very thin subject regions, a subject touching the frame, or a large share of semi-transparent pixels.
Show reviewers the source, transparency grid, light background, and dark background together. Their choices should be approve, send to manual mask, or restore source. Do not hide the original.
A benchmark should sample every catalog class and record defect type, not only a single pass rate. That tells you which SKUs can be automated and which require a different production path.
Next action
Remove one image free, then review the result on both light and dark backgrounds before publishing it.