Changing part of an image used to mean isolating that part first. You drew a selection, feathered the edge, worked inside it, and the quality of the result depended mostly on the quality of your selection. Describing the change in a sentence replaced all of that, and the reason it works is worth understanding.
What the old workflow was really doing
A mask is an instruction about where. It says: change these pixels, leave those alone. Everything about the edit — how the new content blends, whether the lighting matches, whether the edge reads as real — came down to how carefully that boundary was drawn, and then to manual work reconciling the two sides of it.
That is why it took skill. Not because the change itself was hard, but because convincing boundaries are hard.
What replaced it
Instruction-based editing removes the boundary entirely. The model receives the whole source image and your sentence, and regenerates the whole frame conditioned on both. There is no protected region, because nothing is being protected — the model is reproducing what you did not ask to change while rendering what you did.
That sounds more fragile and is usually less so. A mask has to be right at every pixel along its edge. A description only has to be clear.
Why the face survives
The question people ask first is why the face does not drift when everything is being regenerated. The answer is that identity is one of the strongest signals in the source image, and the model is conditioned on all of it. Preserving the face is the path of least resistance; changing it would require the description to push in that direction.
It is not perfect. The further your instruction is from the source — a completely different setting, a radically different pose — the more the model has to invent, and identity can slip. Small, specific changes hold better than sweeping ones, which is a good argument for making several passes instead of one.
Writing instructions that work
- Describe the result, not the operation. "Standing by a window in evening light" beats "change the background and adjust the lighting".
- One change at a time. Three instructions in one sentence usually gets you one and a half of them.
- Be specific about what should be different and silent about what should not. Mentioning something you want kept can cause it to be re-rendered.
- Iterate on the result rather than rewriting from scratch. Each pass starts from a better source.
Where it still struggles
Text inside an image remains unreliable. Precise geometric changes — straightening a specific line, matching an exact colour value — are better done in an ordinary editor. And anything requiring pixel-level fidelity to the original outside the changed area is a poor fit, because the whole frame is being regenerated, and regeneration is never byte-identical.
For everything else, the trade is overwhelmingly in favour of describing it. The work moves from your hands to your sentence, and sentences are faster to redo.
Why the change happened so quickly
Mask-based editing was never only about the tool. It carried a skill requirement, and that requirement was the real barrier — not the price of the software but the hours needed before your results stopped looking cut out.
Instruction-based editing removed the skill, not the software. The gap between a beginner and an expert collapsed to the gap between a vague sentence and a specific one, and that gap takes minutes to close rather than months.
Something was lost with it. Masks give exact control: this pixel, not that one. Descriptions give approximate control over everything at once. For most work the trade is worth it, and for work that genuinely needs exactness, the older approach has not stopped existing.
Working in passes
The instinct is to describe everything at once and expect it all. It is more reliable to go one step at a time, because each pass starts from an image that is already closer to what you want.
Set the scene first, since it constrains everything else. Then the subject, then wardrobe or styling, then lighting last, because lighting is the change most likely to disturb what came before. Four passes with one instruction each beat one pass with four.
The cost is that passes compound. Each one regenerates the entire frame, so detail from the original degrades slightly each time — most visibly in fine texture. Past about five passes it tends to show, which is a practical argument for getting the early steps right rather than correcting them later.
What this means for how you work
The practical shift is that the expensive part of editing moved from execution to description. Time once spent on selections now goes into deciding precisely what you want, and that is a different skill — closer to art direction than to retouching.
It also changes what is worth attempting. Under the old workflow, a small improvement rarely justified the effort of isolating the region, so plenty of images shipped slightly wrong. When a change costs one sentence, the calculation inverts: it becomes cheap to try something, see it fail, and try something else, and the results get better because more variations get tried rather than because any single attempt is better.
