Edit it or generate it again? A decision rule for the frame you nearly have

You have a good frame. The composition works, the light works, and there is one object in the corner that should not be there — or a hand that has gone wrong, or a colour that is off.

Two paths. Re-roll and hope the good parts survive, or fix it in an editor. Both feel like a couple of minutes. One of them frequently is not, and the difference is predictable enough to make a rule out of.

The trap in re-rolling is that a new generation is a new image, not a corrected one. You do not get your frame minus the problem; you get a different frame with different problems, and the thing you liked is often gone.

The question is never 'can I fix this'. It is 'how hard was the part I like to get'.

• If the frame is good because of something structural you can restate — the composition, the lighting setup, the subject — then regenerating will probably reproduce it. Re-roll.
• If it is good because of something you cannot describe — a particular fall of shadow, an expression, a happy accident of texture — then it will not come back. Edit.

Most near-misses are the second case, and most people choose the first. That is the spiral: forty re-rolls chasing a quality that arrived by accident on attempt three.

Edit when the problem is:

• An extra object. Removing something is a solved problem and takes seconds.
• A colour cast or exposure issue. This is grading, not generation. Never re-roll for it.
• A small area of texture or detail. Localised inpainting handles this without touching the rest.
• Anything requiring precision. Type placement, alignment, a crop.

The common thread: the problem has a boundary. You can draw a shape around it, and everything outside that shape stays exactly as it is. That guarantee is the entire advantage of editing and it is worth a great deal.

Regenerate when the problem is:

• Composition. The subject is in the wrong place. No amount of editing relocates it convincingly.
• Lighting direction. The whole frame's shadows are wrong. This is not a fixable region.
• Perspective or scale. Something is the wrong size relative to everything else.
• The frame is boring. A dull image cannot be edited into an interesting one, and the attempt consumes an hour.

These are all decisions the model made early and propagated everywhere. Changing the relevant prompt block and generating again is both faster and better.

For work that has to be right rather than merely good, the reliable method is to generate elements separately and compose them.

Generate the environment empty. Generate the subject on a plain ground. Composite, grade the two together, add type. Each generation is a simpler request, which means a higher hit rate, and each element can be replaced without touching the others.

This is slower to set up and dramatically faster to revise. When a client asks for the product moved slightly, you move a layer instead of regenerating the shot and losing everything else — which is the difference between a five-minute revision and a lost afternoon.

When a near-miss appears, name the problem out loud and place it: bounded or structural, reproducible or accidental. The path follows from that and you act once instead of drifting between both.

The failure mode is doing neither deliberately — a re-roll, then an edit attempt, then more re-rolls, without ever deciding. That is where the afternoons go.

Save the near-miss either way. A frame you rejected for one flaw is often exactly right for a different slot two weeks later.