Fixing part of a picture is a different job from asking for another one

Nearly right is a different situation from nothing yet

The first stage of image work is exploratory. You do not know what you want in any detail, generations are cheap, and each one narrows the field. Wholesale regeneration is the correct tool there, because there is nothing yet worth preserving.

That stage ends abruptly. At some point an image arrives with the right composition, the right light and one wrong element — a hand that does not work, a sign with garbled text, an object in the wrong corner. From that moment the goal has inverted: you are no longer searching, you are protecting.

People keep searching anyway, and the pattern is familiar. Twenty more generations, each with a slightly amended prompt, none of which reproduces the composition that was already good. The frustration is real and self-inflicted, because the request being made is for a new picture rather than a repair.

Local editing exists because the global prompt cannot be surgical

Amending a prompt changes the whole image. There is no phrasing that means keep everything and fix that one thing, because the description applies to the entire canvas and the process runs again from the start. This is a property of how the request is made rather than a limitation to be prompted around.

The tools in this class therefore offer some form of region-limited edit: mark an area, describe what should be there, and leave the rest alone. The names differ and the underlying mechanics differ, but the capability is now standard, and the shift from prompting to selecting is the practical skill worth acquiring.

The most common mistake with region edits is marking too small an area. An edit needs surrounding material to blend against, so a mask drawn tightly around a flawed object gives the process no room and produces a visible seam. Selecting generously and letting the edit reach into stable areas works better.

Save the version you are about to ruin

The discipline that separates a productive session from a frustrating one is almost embarrassingly simple: keep the intermediate files, and keep them with the settings that produced them. The good version is frequently the one from four steps back, and without the file it is unrecoverable.

Where a tool exposes a seed or an equivalent reproducibility handle, record it alongside the prompt. That is what lets you return to the same starting point and vary one element deliberately, rather than hoping the composition recurs. Without it, every return trip is a fresh draw.

The corresponding rule is to branch rather than to continue. When an image is nearly right, make the next attempt from a copy, so a failed edit costs you the attempt and not the picture. Successive edits applied to the same file are how people lose an hour of work to a change they cannot undo.

Some repairs belong in an ordinary editor

Removing a small object, straightening a horizon, adjusting colour, cropping, cleaning up an edge — these are decades-old operations in conventional image software, they are precise, they are reversible, and they do not risk changing anything you did not select. Reaching for generation to perform them is usually the slower path.

Text is the clearest case. Rather than fighting to have a legible word rendered inside a picture, generate the image with a plain area where the text belongs and set the type properly afterwards. The result is sharper, correctly spelled and editable later, which the generated version is not.

A reasonable division is that generation supplies material and a conventional editor assembles it. Composite two generated elements, place the type, correct the colour, and accept that the finished asset was made in the editor rather than requested in one shot. Insisting on a single-prompt result is a constraint you imposed on yourself.

Know when the image is finished or unsalvageable

A useful stopping rule is to compare against the version from several steps back rather than against the previous one. Iteration by small comparisons drifts, and the tenth edit often looks worse than the fourth despite each step seeming like an improvement at the time.

The other judgement is when to abandon. If the same flaw returns after a mask edit, then a wider mask, then a different description, the composition itself is probably fighting you, and further repairs are unlikely to hold. Going back to exploration with what you learned is faster than a twelfth patch.

And there remain pictures this class of tool should not be asked for: an accurate depiction of a real place, a specific product as it actually looks, a diagram whose labels must be correct. Those need a photograph, a rendering or a drawing, and the time spent editing towards them is usually more than the time to obtain the real thing.