AI Video line · stop 13 of 14 · 18 min · members
A short diagnostic order that tells you whether to re-roll, rewrite or change model.
Free with an account
Membership is free: an account opens all 86 script pages. The Lab, Studio Canvas and the paid guides need the $99 pass, paid once. Already signed in on this browser? The page opens by itself.
The default
Because it sometimes works, which is enough to make it a habit.
A generation comes back wrong and the immediate response is to run it again. Occasionally the next one is fine, which reinforces the behaviour — and hides the fact that most failures will not resolve this way.
Re-rolling changes only the seed. It is the correct response to exactly one situation: when the composition and content are right and something specific went wrong by chance.
Every other kind of failure needs a different action, and identifying which you have takes about ten seconds if you know what to look for.
Step one
Check this before anything else, because nothing else will help.
Small readable text, exact brand assets, a specific real person, precise technical or anatomical accuracy, consistent text across a sequence — these are not currently reliable, and no seed, prompt or model change fixes them.
If the failure is one of these, stop generating. The answer is to composite the real asset, set the type separately, or redesign the shot so the requirement disappears.
Ten seconds spent here saves the afternoon that would otherwise go into confirming it.
Step two
One bad image tells you nothing. Four tell you which kind of failure it is.
All four wrong in the same way — the prompt is the problem. The model is doing what it was asked; it was asked wrongly. Rewrite the relevant block.
All four wrong in different ways — the prompt is ambiguous or overloaded. Simplify rather than adding more instruction.
Three fine, one wrong — genuine chance. This is the case where re-rolling is correct.
That single comparison correctly routes most failures, and it is why generating one image at a time makes diagnosis harder than it needs to be.
Step three
You need to know what changed for the result to teach you anything.
When the prompt is at fault, identify which part. Wrong composition is the subject or optics block. Flat and shapeless is lighting. Something appeared that should not have is the constraint block. Wrong overall feel is the treatment.
Change that one block, keep the seed fixed, and regenerate. Now the difference between the two images is attributable.
Rewriting the whole prompt may produce a better image and teaches you nothing, which means the next failure starts from zero again.
Step four
The instinct is to add more instruction. It usually makes things worse.
When results are inconsistent, the cause is often too many competing instructions rather than too few. Every additional clause dilutes the weight on the others.
Strip the prompt back to subject, one lighting instruction and one constraint. Generate. If that works, add back one element at a time until it breaks — you have then found the element that was causing the problem.
This is slower than adding another descriptive sentence and it is the only method that reliably converges.
Step five
It is the largest change and the least diagnostic.
Switching models resets everything — prompt behaviour, vocabulary, settings — so a better result tells you very little about what was wrong.
It is the right move when the failure is a known weakness of the model you are on, established by your own testing rather than by reputation.
Otherwise, work through the prompt first. Most failures attributed to a model turn out to be a prompt asking for something ambiguous, and they follow you to the next model unchanged.
1:1 · two hours