GPT IMAGE 2 in practice：Making Transparent PNGs, from Generation to a Clean Cutout

Codex-generated teaching image. It is opaque, not a H3Max test result or a transparent file.

What checks you use before trusting a transparent PNG? A preview can looks fine, but the downloaded picture may brings a white box into another design. I would like compare a small acceptance checklist with people who works on product pictures or design tools.

I maintain H3Max, so I have a product interest in this question. This is the workflow I am document for its background-removal page. The pictures below is AI-made teaching examples, not test outputs. I do not have benchmark to say this checklist catch every failure.

First, keep the original

If you generate an image, check whether that interface really support transparent output. Saying “transparent background” in the prompt is not enough. JPEG have no alpha; PNG can support transparency but can also be fully opaque. The difference is in the file, not the preview.

If direct transparency is not supported, just keep the original picture and remove its background seperately. You do not needs to generate it again.

Codex-generated opaque source illustration, not a H3Max background-removal result.

Use the full-size original, not a screenshot. Keep the spout, handle and leaves inside of frame. Open the H3Max transparent-background page, select your image, run local background removal, then download the PNG as a new file. This step need no generation prompt.

A quick check before using it

Open the actual downloaded file in an editor, above a separate white layer. Change that lower layer to dark navy. The outside and the handle hole should both shows the new color. The orange body should stay orange.

Codex-generated opaque comparison plate. These are teaching previews, not H3Max outputs or files that prove real alpha transparency.

A checkerboard alone do not prove anything. It may be painted into the image. An alpha channel also does not means the cutout is correct. An external editor can help inspect the alpha and repair the cutout.

Compare with original at 100% zoom. Check leaf tips, the whole spout, the handle rim and the base. If the hole is still gray, clear the leftover background in an external editor. If a leaf is missing, recover that detail from original. Do not paint the hole white.

A pale edge on navy may be old background color, or a real highlight. Check first. An external editor may offer fringe-repair controls; use it carefully so thin details is not removed.

Finish, then check one more time

If there are too much empty space, trim the canvas with a crop tool while keeping the complete subject. If you need another size, prepare the dimensions with a resize tool and keep the proportions.

After either change, reopen the export and repeat the white/navy check. Do not assume alpha survived. Keep original and your checked PNG; leave the test backgrounds out of delivered file.

The part I wants feedback on is acceptance check, not only whether picture looks nice. Is testing on white and navy enough for ordinary product cutouts, or do you also test against textured backgrounds? How you decide whether a pale fringe is unwanted background or a real highlight? If your workflow catch a failure this checklist miss, I would like understand the extra check.