# One change per instruction: a simple rule for editing photos with AI
Most disappointing AI photo edits come from instructions that ask for too much at once. “Make it brighter, remove the person on the left, change the sky and give it a film look” is four jobs in one sentence, and the model has to guess how to weigh them. Splitting the request into single steps is slower on paper but usually faster in practice, because each result tells you what to adjust next.
Contents
A workable order is structure first, atmosphere last. Start with the things that change what is in the frame: remove an object, replace a background, fix a crop. Only after that move to light, colour and mood. If you change the mood first and then remove an object, the removal often repaints the area in a slightly different tone and you end up correcting the mood again. Tools such as nanobanana.org accept plain-language instructions, which makes this step-by-step approach easy to follow.
Keep each instruction concrete and checkable. “Replace the grey wall with a plain white wall, keep the shadows on the floor” can be judged at a glance, while “make it nicer” cannot. Name what must stay the same as clearly as what must change: faces, text on packaging, and logos are the usual casualties when a request is vague.
Finally, save the version you like before trying the next step. A numbered file name such as portrait_v3_wall-white is enough. When a later edit goes wrong you can return to a known good state instead of trying to undo it with another prompt.