AI photo editing has quietly become the more useful half of AI imaging. Not generating pictures from nothing — changing pictures you already have, by describing the change in words.
What it can do now
- Remove things. People in the background, power lines, clutter on a desk, watermarks on your own images.
- Change the background. Same subject, different setting, without a cutout tool.
- Restyle. Photo to illustration, summer to winter, day to night.
- Adjust details. Change a colour, a garment, an expression, while leaving the rest untouched.
- Extend the frame. Fill in beyond the original edges when you need a different crop or aspect ratio.
Why it drifts, and what that means for you
Worth understanding, because it explains every rule below.
When a model edits an image, it is not painting over one region the way you would with a brush. It regenerates the picture with your instruction folded in. Everything is up for renegotiation on every pass.
That is why weaker editors drift. Each edit is a fresh roll of the dice, and small differences compound — three edits in, your subject is a different person wearing the same outfit. The models that hold identity across edits are doing something genuinely harder than producing one attractive image.
How to actually get good results
Change one thing at a time. The single biggest improvement most people can make. "Remove the car and make it sunset and change her jacket to red" produces mush. Three sequential edits produce three clean changes.
Say what to preserve. "Change the shirt to red, keep the face and lighting exactly the same." Models respond well to explicit instruction about what not to touch, and this is where weaker tools give themselves away — you get a subtly different face and cannot say why.
Start with a good photo. AI editing is not rescue software. A sharp, well-lit original edits far better than a blurry one, and no amount of prompting recovers detail that was never captured.
Describe the result, not the process. "The background is a plain grey studio backdrop" works better than "use the lasso tool to cut out the subject." The model is not performing steps; it is producing an outcome.
Check hands, text, and edges. The three places where artifacts hide. Zoom in before you use the result anywhere real.
Keep the original. Always. Every edit is a regeneration, and you will sometimes want to go back two steps.
Compare against the original, not the last step. Drift is invisible step-to-step and obvious end-to-end.
A worked example
Say you have a product photo on a cluttered desk and you need it for a store page.
- "Remove everything on the desk except the product." Check the result — particularly the surface where objects were, which is where filled-in texture goes wrong.
- "Replace the background with a plain light grey studio backdrop, keep the product and its shadow exactly as they are." Naming the shadow matters; without it you often lose the grounding and the product floats.
- "Brighten the overall image slightly, keep colours accurate." Say "colours accurate" explicitly or you may get a warmer, more saturated product than the one you sell.
Three instructions, each checked. Attempting all three at once typically returns a different-looking product, which defeats the purpose.
Prompts that work, by task
Removing something: name it and name its trace. "Remove the person on the left and the shadow they cast."
Background swap: describe the new background and pin the subject. "Studio white background. Keep the subject, their pose, and the lighting on them unchanged."
Colour change: be specific and bound it. "Change only the jacket to deep navy. Everything else identical."
Restyle: name the target medium, not a vibe. "Render as a flat vector illustration with three colours" beats "make it artistic."
Extending the frame: say what should be in the new space. Left to itself, the model invents, and what it invents is usually more of whatever is at the edge.
Where it still falls short
Pixel-exact control. If you need a specific hex value on a specific pixel, or a precise mask, a real editor is still the tool. AI editing is for the 90% of changes where "make it look like this" is a good enough instruction — which turns out to be most of them.
Consistency between runs. The same prompt twice can give different results. Plan to iterate rather than expecting one shot to land.
Fine text and logos. Small type and brand marks degrade. If your image contains either, check them at full size every single time. A slightly wrong logo is worse than no image.
Faithfulness at high edit counts. Each pass compounds small drift. After several edits, compare against the original rather than against the previous step.
Faces you know well. The drift that is invisible on a stranger is obvious on a colleague. Be more careful with images of real, identifiable people — both technically and ethically.
Which model
Editing is a different skill from generation, and models differ sharply at it. The thing to look for is consistency — whether the subject survives the edit unchanged. Google's Nano Banana is currently the strongest on that specific axis.
Being able to try more than one matters here more than in generation, because edit quality varies by subject in ways you cannot predict from a sample gallery.
Common questions
Can AI remove objects from photos? Yes, and this is one of the most reliable things it does — particularly background clutter and unwanted people.
Will it change my subject's face? Weaker models will. The better editing models are specifically good at keeping the subject unchanged, which is why model choice matters more for editing than for generation.
Can I edit the same image multiple times? Yes, and sequential single edits work better than one combined instruction. Just compare against the original periodically.
Does it work on low-quality photos? It works, but poorly. Editing does not recover detail that was never in the original.
Is it better than Photoshop? Different. Faster for "make it look like this", worse for precise control.
Can I use edited images commercially? Usually, but the terms differ by provider and by whether the source image was yours. Check before publishing anything client-facing.
Why does the lighting change when I only asked about one object? Because the whole image is regenerated on every edit. Explicitly pinning the lighting in your instruction is the fix.
How do I keep a character consistent across many images? Use an editing model known for identity consistency, edit from one source image rather than generating fresh each time, and change one attribute per pass.