Six product types where AI background removal still needs a human check

Most cutout tools do the same thing: they predict one mask and paint white behind it. On a clean shoe or a mug that is already close to correct. On the six product types below the mask has to answer a question the model cannot answer for you, and that is why a 98% accurate cutout can still fail a marketplace review.

1. Glass and transparent packaging

Bottles, jars, and blister packs have two edges: the outline and the light passing through the material. A mask keeps the outline and throws away the transmission, so the object reads as a flat sticker. Nothing inside a cutout tool can rebuild what was behind the glass, so these need a deliberate decision — either keep a small refracted strip of the original, or accept a stylised shape and say so in the listing.

2. Hair, fur, and fringed textiles

This is where edge resolution matters more than the model. A single-pixel error is invisible on a bottle and obvious on a fur collar, because the fringe is the shape. Judge these at 200% on a dark background: halos that are invisible on white show up immediately there.

3. Reflective metal and glossy ceramics

A chrome kettle or a glazed bowl carries the background in its reflections. Replace the background and those reflections become a photograph of the old room. If the reflections already contain the studio, the honest fix is a reshoot; if they contain a grey sweep, matching the new white to the reflected white is enough.

4. Mesh, lace, and netting

The holes are product, not background. Coarse mesh survives a mask reasonably well; fine netting usually turns into a blur, and the transparent pixels inside the weave get filled with white, which changes the product's apparent density. Check a zoomed crop, not the thumbnail.

5. Thin straps, wire handles, and loose threads

Every one of them is one or two pixels wide at catalogue size. A morphological cleanup pass — the kind that removes stray specks — will also remove a strap. Compare the result against the original at full size before you trust the preview.

6. White and near-white products

A white candle on a white background loses its silhouette. The cutout can be perfect and the listing still unreadable. Keep the natural contact shadow, or add a one-pixel inset border, so the object's outline survives.

What actually helps

mode        what it does                              when to use it
Standard    single-pass mask, fast                   clean edges, solid objects
Quality     generative pass, up to 2048px             overlaid text, matted texture

Quality mode is the one that can remove printed text or a watermark that bleeds into the object, and it is also the one that can quietly repaint a label. Treat every Quality result as a new photograph: put it beside the original and check the label, the colour block, and any printed measurement before publishing.

A five-minute check before you publish

1. Sample the background at five points — corners, mid-edges, centre. Every one should read the same value.
2. View the edge at 200% on a dark background, then at 100% on white.
3. Compare the object's colour against the original on a calibrated screen, not the preview thumbnail.
4. Check the silhouette: straps, handles, threads, and any part thinner than the eyelash rule.
5. Confirm the file is the size and format the marketplace asks for — transparent PNG where the background should not be white.

The tool I use for this is a browser-based cutout that takes JPG, PNG or WebP up to 10MB, offers a Standard and a Quality mode, and can output transparent PNG; previews are free and uploads are deleted within two hours. Whatever you use, the lesson is the same: the cutout is a proposal, and the six categories above are where you have to accept or reject it.

I am the developer of CleanWhiteBG (https://cleanwhitebg.com), the tool linked above, and this post is written from that work.