What "editing" means now
Classic image editing works with tools: you select, you mask, you push layers around. AI editing works with an instruction instead: "take the bin out of the background", "make it evening", "change the shirt to blue". The model gets the original image and the sentence, and hands back a new image.
The important thing to grasp is that it does not repair, it regenerates. There is no such thing as "it will only touch that part". The model rebuilds the entire image while trying to follow the original. Every oddity comes from that: it is why the shade of the background, the light, or a small feature of a face sometimes changes even when all you asked for was one button in blue.
The two sentences that matter most
The largest provider's documentation recommends one specific phrasing, and experience says it is the single most useful trick in this whole area.
GOOD:
Change only the color of the mug to dark green.
Keep everything else exactly the same: the lighting, the background,
the position of the objects, the texture of the table.
BAD:
Make the mug green.
The point is the second sentence: you list what must stay unchanged. And — this is the non-obvious part — the keep list has to be repeated in every round, not stated once at the start. The provider adds that itself, along with the admission that even this is not perfect: repeated editing can still alter details you wanted preserved.
The example is in English on purpose. Image models are trained overwhelmingly on English captions, so if you work in another language it is worth writing the instruction in English even when the picture is for a local audience. The structure is what carries over: you list what must not change. Where that gap comes from is covered by models and smaller languages, with measurements.
Why the image degrades round by round
Several measurements in 2026 shone a light on this, and the results agree to a surprising degree.
- From the third round onwards the subject starts to distort measurably: proportions slide, outlines soften.
- Up to about five rounds the result stays usable in most cases.
- Above ten rounds the damage is irreversible — not a small flaw, but the image becoming a different image.
- The decline is not even: removing an object works well in a single round, while across several rounds the success rate falls to roughly half, and counting the things in the picture degrades more steeply still.
The fidelity dial
Some providers give you a separate setting for how tightly the model should stick to the original image. Where it exists, it is worth turning up for any job that starts from a real photograph — people or products, for instance.
Worth knowing, though: this setting is neither standard nor permanent. One model generation had it, the next switched it off, and the newest has no documentation about it at all. Do not build a workflow on it; treat it as a bonus when it happens to be there.
When to use a mask
Conversational editing is convenient, but it is not always the better choice. If you know exactly which part of the image you want swapped, select it — most tools let you paint a mask with a brush, and then the model only rewrites the selected area.
The difference is practical rather than stylistic: with a mask the rest of the image is guaranteed untouched, so there is no drift however many rounds you run. Mask-based operations — removing a blemish, extending the frame, swapping the background — form a family of their own, and inpainting and background replacement covers all three at once. Changing the composition after the fact is not editing at all, though: if the arrangement is the problem, describing the layout offers a far more workable approach.
Common misconceptions
- "I will just polish it bit by bit until it is perfect." This is the most expensive myth of the lot. After five rounds it typically gets worse, not better. If it is not there yet, start again from the original with different wording.
- "The original file is still there, so there is no risk." The file, yes; your work, no. A flaw spotted in the seventh round means seven rounds of work go in the bin.
- "If I do not mention it, it will not touch it." Exactly backwards. Whatever you do not explicitly protect, the model is free to reinterpret.
- "A bigger model does not drift." Less, but it does. The effect comes from the mechanism, not from the size of the model.
What to take away
AI image editing is good when you use it in few rounds, precisely worded: say what should change, and list what should stay — in every single round. If the third round still is not what you wanted, stop polishing: go back to the original, or mask the part you are actually talking about.