There is a strange moment in almost every AI video experiment.
The first generation feels impressive. A still image moves. A camera drifts through a scene that did not exist a minute ago. Lighting changes, fabric moves, and suddenly an idea that was sitting in a notes app looks almost like footage.
Then you try to make the second clip.
That is usually where the real work begins.
I have been thinking about this while exploring Seedance 2.5 and similar video workflows. The interesting question is becoming less about whether a model can generate motion at all. What matters more is whether you can carry an idea through several attempts without losing the parts that made the first one useful.
A good result creates a new problem
Suppose you are working from an image of a small café at night.
The first result gets the atmosphere right: warm windows, wet pavement, slow camera movement. But the person near the entrance moves strangely.
You generate again.
Now the movement is better, but the lighting has changed.
Another attempt fixes the lighting but alters the composition.
None of these results is necessarily bad. The problem is that generation and direction are two different things.
Once you find something worth keeping, you need a way to communicate:
keep this composition;
preserve this visual reference;
change the movement;
make the next shot feel related to the previous one.
That is much closer to directing than simply prompting.
References are becoming part of the creative language
This is why reference-based workflows are more interesting to me than increasingly elaborate prompts.
A prompt can describe what you want. A reference can show what you mean.
That distinction matters when working with a particular object, environment, color palette, camera feeling, or sequence of shots. Instead of packing every requirement into another paragraph of text, creators can begin treating references as part of the instructions.
The workflow around Seedance 2.5 is useful to examine from this perspective: text or images can establish an initial direction, while additional reference material can help shape subsequent generations.
But references do not magically create continuity. Someone still has to decide what should remain stable and what is allowed to change.
I started keeping a tiny “continuity note”
Nothing sophisticated. For each useful generation, I record four things:
Keep: the elements that already work.
Change: the specific problem in the current clip.
Reference: the asset that best communicates the intended look or motion.
Ignore: details that do not need to survive into another version.
This has made experimentation noticeably less chaotic.
Instead of writing a completely new prompt every time, the next generation has a purpose.
That sounds obvious, but generative tools make it surprisingly easy to confuse making more versions with making progress.
The first clip is becoming the sketch
Perhaps that is the larger change happening in AI video.
The first result does not have to be the finished result.
It can be a moving sketch.
You can use it to discover that the camera should be lower, that an action needs more time, that the background is too busy, or that the scene works better without the element you originally thought was essential.
For creators, designers, and small teams, that may ultimately be more useful than chasing a perfect one-shot generation.
The interesting future of AI video may not be a button that makes a finished film.
It may be a workspace where an unfinished visual idea can gradually become more specific.
And that changes the question from “What can this model generate?” to something much more practical:
“Can I keep directing after the first generation?”