The most expensive habit in AI image work is regenerating a whole picture to fix one part of it. A cable in the corner, a grey backdrop that should be white, a person in the background of an otherwise good shot — none of those need a new render. They need a cleanup tool.
Generation and cleanup are different operations
A generator makes a new image every time you run it. Even with the same prompt and the same references, the lighting shifts, the product moves, the details change. That is fine when you want options and ruinous when you had the right shot already.
A cleanup tool works on the picture you have. The subject stays exactly as photographed; only the part you targeted changes. For anything involving a real product, that distinction is the difference between a listing that matches the item and one that does not.
The four tools, and what each is for
| Tool | Use it for | Keeps the subject? |
|---|---|---|
| Background remover | Isolating a subject from its setting | Yes, pixel for pixel |
| Transparent background | PNG cutouts for layouts and overlays | Yes |
| White background | Marketplace listings that demand plain white | Yes |
| Change background colour | Brand-coloured backdrops for a set | Yes |
| Remove object | Deleting something inside the frame | Everything except the area you mark |
The first four are one-click: upload, download. Object removal is the one that asks for input, because the tool has to know what you consider unwanted.
Removing an object well
Object removal works by rebuilding the area you mark, using the surrounding pixels as context. Two things follow from that.
First, mark generously. A tight outline around a wine glass leaves its shadow and reflection behind, which reads as a ghost. Include the shadow, the reflection and a little of the surface underneath.
Second, simple backgrounds rebuild better than busy ones. Removing a cable from a plain wall is nearly invisible. Removing a person from a patterned rug in front of a bookshelf asks the tool to invent a lot of structure, and invented structure is where artefacts appear.
When removal is the wrong call
If the thing you want gone occupies a third of the frame, you are not removing an object, you are asking for a new background. That is a job for image to image with the subject as a reference — or for a background swap, if the subject can be cut cleanly.
Backgrounds: cut, replace, or recolour
Three different problems get called "background" and they have three different answers.
Cut it out. You need the subject on transparency to drop into a layout. That is the transparent background tool, and the output is a PNG with a real alpha channel.
Make it plain. Marketplaces frequently require pure white. White background does exactly that in one step, and the result is consistent across a set of photos — which matters more than it sounds, because inconsistent whites look like sloppy photography.
Put it somewhere new. This is a generation job, not a cleanup job. Cut the subject out first, then use image to image with that cutout as a reference and describe the new setting. Doing it in that order keeps the product intact; describing product and setting together in one prompt usually redraws both.
What cleanup costs
The cleanup tools are credit-cheap compared with a full render, which is the point. Removing a background runs a few credits; object removal a little more because it has to rebuild the area. Compare that with 8 to 20 credits for a fresh render on a high-resolution model, which also gives you a different picture than the one you were trying to save.
The file tools cost nothing at all. Compress, resize, convert and rotate run in the browser, so trimming a 4K export down to a web-sized JPEG is free. Reaching for a generator to shrink a file is a surprisingly common and entirely avoidable expense.
An order of operations
Cleanup before enhancement, always. Remove the distraction first, then run any generative pass on the cleaned image. Doing it the other way round means the model spends effort rendering detail into something you are about to delete.
Within a set, do the same operation to every image in one sitting. Backgrounds removed at different times with different settings drift, and drift across a product set is visible even when each individual image looks fine.
Export last. Keep the working copy at full size; produce the platform-sized files at the end, once nothing else will change.
A quick decision tree
Is the subject correct in the picture you already have? If yes, never regenerate — clean up instead. Is the unwanted thing behind the subject, or inside the frame? Behind means a background tool; inside means object removal. Does the subject need to move, turn or change material? That is a real generation job, and image to image with references is where it belongs.
Everything else — size, format, orientation, weight — is a file operation and should cost you nothing. The workspace keeps the generators and the tools in the same place, so moving between them does not mean re-uploading anything, and plans covers what the generative steps draw from your balance.
Marking the area: what a mask is doing
When you mark a region for removal, you are telling the tool two things: what to delete, and what it may use as reference for rebuilding. The second part is why mask shape matters as much as mask size.
A mask that hugs an object gives the tool almost no surrounding context to sample from, and rebuilt pixels look pasted in. A mask with a margin — including the object's shadow and a little clean surface — gives it a clear picture of what the area should look like when the object is gone.
Marking several small things separately usually beats one large mask that swallows the space between them, because each rebuild has more real pixels around it.
Keeping a product set consistent
Eight listing photos cleaned on eight different days will not match. Backgrounds land at slightly different whites, edges are cut with different tightness, and the set reads as amateur even though each image passes on its own.
Process a set in one sitting, in the same order, with the same tool. If one image needs object removal and the others do not, do the removal first, then run every image through the background step together.
What cleanup cannot fix
| Problem | Cleanup? | What to do instead |
|---|---|---|
| Distraction in the corner | Yes | Object removal |
| Busy background behind the subject | Yes | Cut out, then place on a new backdrop |
| Subject is blurry or badly lit | No | Re-shoot, or regenerate from references |
| Wrong angle of the product | No | Image to image with references |
| Label text unreadable | No | Supply the label as a reference image |
The pattern: cleanup edits what is in the picture, generation decides what the picture is. Problems of the second kind never get solved by the first kind of tool, no matter how many passes you run.