AI photo editing goes wrong when every change happens at once. Cleaner frames come from locking the original first: exposure, white balance, and a short note about what cannot move. Then ask the model for one change, such as lifting a dark doorway or clearing a cable, and compare the result at full size.
Name the subject, the light, and the edge of the edit. "Keep the face, hands, and jacket. Soften only the street behind them" is easier to accept or reject than a vague request to improve the shot. Look for waxy skin, bent window frames, and letters that stop making sense.
Keep the source file. If a pass is too smooth, reduce the strength or put the original grain back on top. When the same wording works twice, save it and reuse it across the set.
For a short note on the image model used in this loop, see GPT Image 2.