Remove a price sticker from a product photo and check the label

Remove a sticker from a product photo with Ideogram 4.5 on Sume, then check the label text underneath. A small mask on GPT Image 2.5 limits the edit.

5 min readSume
All posts

To remove a price sticker, describe the sticker by position and say what is under it: "Remove the round price sticker on the bottom right of the box. Restore the printed label under it, keep every other word unchanged." Send the photo to Ideogram 4.5 as the only reference. The risk is the text under the sticker: the model must invent it, so read the result and, if the words are known, replace that patch with the real artwork.

Two routes

Route one is the prompt above, at $0.0375 per low-quality attempt on Sume (the provider list of $0.03 times 1.25). Morphic's page lists "Remove the cables and the power strip, keep the desk" as an example of the remove-and-keep shape, so the wording is close to what the vendor describes.

Route two adds a mask. The Image API docs document mask_url for openai/gpt-image-2.5, so a small mask around the sticker limits where the model may change pixels. The docs do not say which colour marks the edit area; render one test and look. GPT Image 2.5 is token-priced ($30 per million output tokens), so at 1024x1024 xhigh output alone is $0.09366 and max $0.21072, before input tokens, so keep the quality at the tier you need.

The text under the sticker

Test with the worst case first. Pick the photo where the sticker covers the most print, run three low attempts, and read each result. If the model gets that one right, the easy photos will be fine; if not, you know before you have run the batch that the paste-back route is needed. Three attempts cost $0.1125 at the low tier on Sume.

Stickers hide printed words, and a model that has never seen them will produce plausible letters. For a barcode, an ingredient list or a legal line, plausible is wrong. Take the clean artwork from the pack's design file, place that patch over the edited area with Pillow, and keep the model's work for the surrounding surface and shadow. The barcode guide covers the paste-back.

Checklist

Clean-up on resale or catalog photos often means removing more than one mark. Do them one at a time: a sticker, then a scuff, then a glare spot, each as its own pass. A single prompt that lists three removals asks the model to decide what counts as damage, and it may remove a feature of the product, such as an embossed logo, along with the scuff.

If the sticker curls over an edge, say so. A sticker wrapping a corner is part shadow and part adhesive, and a prompt that ignores the corner leaves a ghost of it behind.

Sticker removal checks by route (read 2026-10-07)
CheckPrompt only (Ideogram 4.5)Mask (GPT Image 2.5)
Sticker goneLook at 100% zoomLook at 100% zoom
Text under it correctRead it; replace if knownRead it; replace if known
Rest of the box unchangedRun a pixel diffMask area only, diff to confirm
Per attempt on Sume$0.0375 lowToken priced; read the endpoint

Sources

Related posts

More in Use cases

All Use cases posts

Written by Sume