Continuity is the wall: what has to change before AI can cut a whole ad
This update was drafted on a schedule by the AI I build with, from real project notes — part of the vibecoding experiment this blog documents.
I said in the last post about the video pipeline that continuity is the one limit to watch — the day a model can hold the same product across a cut, the shape of the work changes again. I want to actually take that apart, because "continuity" gets used as one word for about six different problems, and they're not all equally hard, and knowing which is which tells you roughly how far off this actually is.
When I run Cadence and generate two clips of the same can, I get two cans. Cousins. Not the same can. And "the label moved a little" sounds cosmetic until you're trying to cut between three angles of a product in a 20-second spot and the product is visibly a different object in each shot. That's the wall. Here's what's actually behind it.
Continuity is really six problems
Object and character identity. The same can, the same face, the same logo, shot to shot. This is the headline one and it's the furthest along — reference images, first-and-last-frame conditioning, character models trained on one face. It half-works. "Half" is not good enough for a paying client, but it's the direction with the most movement.
Set and environment persistence. Cut away from the room and cut back, and it should be the same room — same furniture, same window, same clutter in the same places. Models don't hold this at all right now. Every generation rebuilds the world from scratch.
Lighting and color continuity. The key light stays camera-left, the grade stays warm, the time of day doesn't jump. An editor fixes small versions of this in the grade. Big versions — the sun moving between shots that are meant to be seconds apart — are unfixable in post and the model gives them to you constantly.
Wardrobe and props. The watch is on the same wrist. The jacket is the same jacket. Small, but audiences catch it instantly because we're wired to.
Spatial geometry. If she's on the left looking right in shot A, she's still on the left in shot B. The 180-degree line. The model has no concept of a persistent stage, so screen direction is a coin flip every generation.
Motion carrying across the cut. A hand reaching at the end of one shot is mid-reach at the start of the next. This is match-on-action, the basic grammar of editing, and it needs two generations to agree about a movement neither one knows the other is making.
Notice that only the first of these is getting real attention. The rest are barely acknowledged as problems yet.
Why re-rolling doesn't get you there
The reason all six fail has one root. Each generation is an independent sample. The model isn't editing a scene that exists — it's drawing a fresh one from a distribution every time you hit go, and nothing carries over because there's no thing for it to carry over from. That's also why you can't revise by note: "same shot, slower" gets you a different shot that's slower, because there was never a persistent shot to modify.
You can bolt on patches. Feed a reference frame. Condition the last frame of clip one as the first frame of clip two. Use video-to-video so an existing clip constrains the new one. Build a character model so at least the face is stable. These help, and Cadence uses them. But they're all attempts to smuggle state between two processes that are fundamentally stateless, and it shows — the seams are visible, the drift creeps back over a few seconds, and the more you constrain the model the more of its quality you give up.
What would actually have to change
The real fix isn't a bigger model. It's a different object in the middle. Something persistent that the generation edits rather than re-samples — a scene representation that holds the can, the room, the light, the blocking, and survives between shots, so that "now show me the wide" is a camera move within a world that already exists instead of a brand-new roll of the dice.
You can see people reaching for this from a few directions — 3D and gaussian-splat intermediates you can re-shoot from any angle, world models that maintain state over time, hybrid pipelines where the AI dresses a scene that's structurally locked. None of it is production-ready and I don't know which approach wins. But that's the category of change it takes. Not "the shots got prettier." "There's now a persistent thing the shots are of."
How you'll know it landed
Here's the test I'll use. Generate a shot. Then, without regenerating it, ask for the reverse angle of the same moment — same product, same room, same light, camera on the other side. If what comes back is recognizably the same scene from a new position, the wall's down and a lot of what I've written about this is obsolete. Until that specific thing works, every prompt-to-video tool is a shot generator, and the ad still gets cut by a person.
I don't have client work to point at — four reels rendered, the studio's live, nobody's paid me for motion yet. So this is a report from the build. But the gap is specific and I think it's structural, not a matter of scale, and it's worth watching the right part of it: not whether the clips look good, they already do, but whether anything persists between them.
This one's auto-drafted from my notes on a schedule. If a number isn't in the notes, it doesn't show up here — I'd rather leave a blank than make something up.