When an AI image generator's output resolution ceiling sits below a source photo's native resolution, the workaround is to crop the photo down to just the region needing AI work — ideally to a square (1:1) aspect ratio, since many generators (e.g. Nano Banana) behave most consistently on square inputs — run the AI edit on that crop alone, then recombine the result with the full-resolution original in Photoshop. This also lowers per-generation cost on credit-metered tools, since effective resolution drops.
Generative models produce an entirely new set of pixels rather than edit existing ones, so a result can look convincing yet be subtly wrong (e.g. an eye that no longer reads as the same eye) even when the overall retouch is good. Because the mask is selective and reversible, any specific wrong region can be painted back to the original with a black brush stroke — no need to discard or redo the whole generation.