Last month I spent forty minutes retouching a photo of a coffee cup for a cafe’s menu page. Erasing a crumpled napkin, rebuilding the table edge behind it, cleaning a green reflection out of the glaze. Then I opened an AI image editor online, dragged a rough box over the napkin, typed “remove napkin, continue the wood grain,” and got a cleaner result in about eleven seconds.
That’s the honest pitch for these tools. Not magic, not the death of photographers, just a very fast way to handle the tedious part of image work. The trouble is that the gap between the demo video and your fourth failed attempt is wide enough to eat an entire afternoon. So here’s a practical map of what these editors do well, where they quietly break, and a workflow that has held up under deadline.
Editor or generator? The distinction matters more than you think
A generator builds an image from a text description. An editor changes an image you already have. Most modern tools blur that line, which is exactly why people get frustrated: they upload a photo, type a wish, and expect the entire scene to rearrange itself around it.
The editing functions that are genuinely reliable right now are narrow and specific:
- Inpainting, where you paint over a region and describe what should replace it
- Object removal, which is inpainting with the surrounding background reconstructed for you
- Background isolation, cutting a subject out cleanly enough for a product page
- Generative expand, stretching the canvas and filling the new space
- Upscaling and denoising, doubling resolution without turning skin into plastic
If your task fits one of those five, you’ll probably be happy with the result. If your task is really “make this photo look like it was shot by someone else entirely,” you want a generator instead, and that’s a separate skill worth learning if you want to master the art of the AI image creator.
Where the results actually hold up
Removing clutter from real photos
Power lines, exit signs, a stranger’s shoulder in the corner of a wedding shot. Select with 10 to 15 percent extra margin around the object, keep the prompt blunt (something like “plain overcast sky, same grain”), and run it twice. The second pass usually cleans up the smear the first one left behind.
Swapping backgrounds on product shots
Cutouts are close to solved. Matching light is not. A watch pasted onto marble will float unless the new background has a contact shadow matching the original light direction, and the reflection runs the right way. Nine times out of ten the fix is a two-minute gradient shadow you paint in yourself, not another generation.
Rescuing low-resolution images
Doubling a 900-pixel product photo for a landing page works well. Pushing a 400-pixel screenshot up to print size does not: the model invents detail, and invented detail in faces and fine type looks uncanny at 100% zoom. Upscale to what you need, then stop.
Prompting inside an editor is a different language
When you’re generating, you describe a scene. When you’re inpainting, you describe material, light and texture for a small patch of image. “A beautiful forest background” gives you a stock-photo mess. “Dappled shadow, cool green fill light from the left, shallow depth of field” gives you something that sits down and behaves. That habit of describing light rather than subjects is the same discipline behind ChatGPT image prompts that actually work, and it transfers to editors almost word for word.
The free tier is a test drive, not a workspace
Almost every tool will let you edit a handful of images before the walls go up: a watermark in the corner, exports capped at 1024 pixels, credits burned on every retry, and a commercial licence buried three clicks deep in the FAQ.
Retouching is iterative, so credits vanish faster than you expect. Twenty failed inpaints is completely normal for a tricky removal. It’s worth reading a comparison before you commit, and these free AI image tools that are actually worth your time cover which ones keep the watermark off and which quietly claim rights to your uploads. Same due diligence applies to editors.
A 20-minute workflow for one product photo
This is roughly what I do now, and it’s cut my retouch time on simple e-commerce shots by about two thirds:
- Duplicate the original and work on the copy. Never on the file you might need to fall back to.
- Do geometry first: straighten, crop, fix perspective. AI fills behave badly on a skewed canvas.
- Remove obvious clutter in one pass, using small selections rather than one giant mask.
- Swap the background last, then paint contact shadows by hand.
- Upscale once, at the end, after you’ve stopped editing. Re-editing an upscaled file drags old artifacts back in.
- Export at the largest size you’ll plausibly need and archive the layered version.
Things these editors still get wrong
Hands, for one. Fine type on packaging, for another: ask for a label and you’ll get convincing gibberish. Repeating patterns like brickwork or floor tiles drift out of alignment across a fill. Two objects in the same scene often end up lit from different directions, which your eye catches instantly even when you can’t name what’s off.
There’s a rights question too. Output from a model trained on scraped images sits in murky territory in several jurisdictions, so for paid client work it’s worth reading the actual terms rather than assuming. Screen every export at full size before it goes anywhere near a client.
Where the edited image ends up matters
Half the images I fix land in a slide deck or a pitch document, and that’s where a good retouch can still look cheap: dropped into a 4:3 placeholder at the wrong crop, squashed to fit. If you build decks regularly, a template-driven builder saves you from re-laying-out every image by hand, which is roughly why I’d never build a presentation from a blank slide again.
What still decides the result: your eye
The tool will cheerfully give you twelve variations of a background. It won’t tell you the shadow is wrong, the subject sits too far left, or the colour clashes with the brand palette. Those calls are still yours, and they’re what separates a retouch that looks professional from one that looks generated.
So pick one editor, learn its inpainting quirks, keep your originals somewhere safe, and check every export at real size before it ships. The eleven-second coffee cup is real. The forty minutes it replaced were mostly the software’s fault, not mine.