Where style actually lives in a pipeline
The stage where you can most easily change a look is rarely the stage where the look is actually decided. Knowing the difference is most of the job.
When output looks wrong, there is a strong pull toward fixing it at the stage you have the most control over. That is usually the prompt or the conditioning, because those are right there and iterating on them is fast.
It is also frequently the wrong place, and the resulting hours are the most common waste I see in this work — including my own.
The eye tells you that, not where
A trained eye is very fast at detecting that something is off. It is no help at all in locating the cause. Those are separate skills and they get conflated constantly, including by people hiring for one and expecting the other.
Take a clip whose look is inconsistent. Plausible causes:
- The style definition is genuinely inconsistent.
- The style is fine, and the temporal pass is destroying it.
- The style is fine, and each segment is conditioned independently.
- Reference imagery is internally inconsistent, so the target is an average.
- Style conditioning is strong enough to override content.
- Output normalisation is running per segment.
Six explanations, one symptom. No amount of looking harder distinguishes them.
Isolate stages, don’t tune them
The move that works is boring: hold everything constant and change one stage at a time, including stages you assume are fine. Not iterating on the stage you suspect — testing stages you don’t.
Concretely, on the list above: run two segments with identical conditioning and compare them. If they disagree, it is not the style definition. Then disable the temporal pass and compare again. Each test eliminates a whole class of cause, and three or four of them will usually land on the real one.
This is unglamorous and it is faster than any amount of adjustment.
Where the leverage tends to be
A pattern I now expect: the stage with the most influence over the final look is often not the generation step. It is something upstream and unfashionable — what got anchored, what the reference set contained, where a boundary was placed.
Those stages are easy to overlook because they seem like plumbing. But the generation step is the stage you control least, and anything you can decide before it runs, you decide with certainty rather than by suggestion.
So the first question on an unfamiliar pipeline is not “how do I make this look better.” It is “where in here do I actually have authority?” — and then working out how much of the result can be settled there.
If any of this is relevant to something you're building —get in touch.