How to remove text, watermarks and objects from product photos: AI inpainting guide

What inpainting actually does, and what it cannot do

Inpainting does not recover the original pixels under a mark, because those pixels were never in the file. A watermark, caption or sticker replaces the scene underneath it, so the tool must invent a plausible replacement: continuing a wall, repeating a fabric weave, extending a shadow, or blending a sky gradient across the gap. That is why the background matters more than the tool. A logo over a plain wall, a caption over a flat sky, or a cable over a wooden table is easy, because the surroundings give a clear pattern to continue. The same logo over a face, over small printed text on the product, or over a pattern with no clean repeat is hard, because there is no single plausible answer, and the model will guess.

Two limits matter before you start. First, removing a visible mark does not remove invisible provenance signals: content credentials and statistical watermarks work separately from the visible logo, and a clean-looking edit can still carry them. Second, a blur is not inpainting done badly, it is a different job: blurring hides a mark, inpainting rebuilds the scene. If your result looks like a smudge, do not ship it hoping the thumbnail hides it. Shoppers zoom, and a smudge on the hero image reads as a damaged listing. Regenerate with a tighter brush, or use the fallback in the table below.

The legal line, stated plainly

Only remove marks from photos you own or have the right to edit: your own product shots, your own old watermark or signature, and supplier photos where your licence allows editing for your listing. It does not cover stripping a stock-photo watermark to use an unpaid image, removing another photographer's credit to pass their work off as yours, or erasing a brand logo from a product you do not sell so the photo misleads a buyer. If unsure, ask the rights holder, buy the clean file, or shoot the product yourself.

The same honesty rule applies after the edit. A cleaned photo must still show the product you actually ship, in its true colour, shape, material and condition. Removing a scratch, a dent or a stain from a used item, or erasing a warning label, is not cleanup, it is misrepresentation, and it produces returns and disputes. Clean the background and the stray marks around the product. Leave the product itself alone, except for normal colour and exposure correction, and always work on a copy of the original.

The five-step clean edit

Work on a copy at full resolution, never on the only original, and start with the smallest job first: a clean corner edit shows how the tool behaves on this photo before you attempt fabric or pattern. Brush only the mark, plus a small margin of a few pixels at full size. A brush far too large forces the model to rebuild good pixels as well, which turns clean areas waxy. A brush too tight leaves a halo of the old mark at the edge, which returns as a ghost outline in the thumbnail.

Describe the fill only when the tool asks for it, and describe the background, not the product: name the surface and the light, as in the examples below. Then compare, do not just accept. Flip between original and edit, zoom to 100 percent, and check edges, grain and shadows. Grain is the giveaway most people miss: a rebuilt area smoother than the surrounding photo looks like a sticker even when the colour matches. If the grain is wrong, regenerate once with a slightly larger brush; if it is still wrong, stop and use the table fallback. Two failed attempts usually mean the background is too complex, not that a third will be lucky.

Export only after the check passes, at the resolution your channel needs, at least 2000 pixels on the longest side for a marketplace hero, and keep the untouched original alongside the edited file. The erase tool handles the brush-and-fill step in the browser, and the remove-background guide covers the next step: turning a cleaned photo into a cutout on a pure background.

What to do for each kind of mark

The date-stamp row is the one sellers underestimate. On a plain corner it is trivial; over a patterned border, a tiled floor or a printed box edge, the same stamp becomes a pattern-matching problem, and the rebuilt border shows a kink the eye catches instantly. The rule is simple: if you can describe the background in one sentence, inpaint it. If you need a diagram, the background is too complex, and cropping, a background swap or a reshoot will be faster and more honest.

The pre-publish check, in order

Run this check on every cleaned photo, not just the difficult ones, because the easy ones are where a halo or grain mismatch slips through. First, the thumbnail: does anything draw the eye to the edited spot? Second, full size: does the area match the surrounding colour, grain and shadow direction? Third, at 100 percent, trace the brush edge slowly: any hard line, repeated texture or ghost of the old mark fails the photo. Fourth, the product again: colour, logo, label and proportions unchanged, no invented detail, no removed flaw. Fifth, the file: correct size and format, original saved separately.

For a batch, use two passes: a fast thumbnail pass over every image, which catches halos and colour shifts, and a slow 100 percent pass over each hero image, which catches grain and edge problems where they cost most. The small-shop photo guide shows how a cleaned photo becomes a full listing set. A photo that fails either pass is not a near miss: regenerate it once, and if it still fails, crop, swap the background or reshoot it.