I Directed a Donut Commercial With AI. Then Physics Got Involved.
I built a full :30 broadcast spot for a campaign that doesn't exist — Dunkin' x Smucker's, self-initiated, no client, no brief. Twelve shots. Two women, a sidewalk, a jelly-filled donut, one dramatic collision.
Here's what nobody tells you about directing AI video: the tools don't know what a donut is supposed to do.
The collision shot went first — and it never worked. Not once.
Two people bumping into each other, a donut and a jar going airborne — it should be simple. It is not simple.
I ran it over and over. Sometimes the women didn't walk toward each other correctly. Sometimes they avoided each other entirely. Sometimes the camera ignored the direction I'd given it. And sometimes something completely unrelated went spectacularly wrong in the background.
So I did the thing an actual director does when the shot they wanted doesn't exist: I built around it. I cut to just before impact and just after, letting the motion in the surrounding shots imply the collision instead of showing it.
It works. Watching the spot, your brain fills in the middle beat and the sequence reads fine.
But I know what's missing. I wanted that shot. I still do. There's a version of this commercial in my head with a real collision in it, and it's not the version that exists.
The jelly jar came next, and it tried to kill me too.
I needed one thing: a jar of Smucker's, falling, spinning, readable label. That's it. That's the whole ask.
Round one: the jar didn't fall. It turned too fast, then too slow, then — I still don't fully understand this — it stopped mid-air, did a little dance, and hung there like it forgot what gravity was.
Round two: fine, it's tumbling now. Great. Except the label had turned into Simlish. Not blurry, not almost-right — fully invented gibberish, confidently rendered like it meant something.
So I fixed the label in Photoshop. Re-ran the prompt. Now it's tumbling correctly, the label is legible, real words, real brand — upside down. On a jar that was otherwise doing everything right.
I was about a second and a half from deciding I didn't actually need a commercial for a campaign that isn't real. But I got it. Eventually.
Then that GD donut...
Every. Single. One of the twelve shots had its own version of this fight. I generated and art directed the imagery shot by shot, directed the motion for each one, then edited the pieces into the finished :30 and cut that down to :15.
Nothing about "AI makes this fast" was true in the moment. It was fast the way running down twelve different hallways is fast — a lot of motion, most of it wrong, until one door opened. Or, in one case, never opened at all.
Here's the actual lesson, once I stopped laughing:
AI didn't direct this commercial. I did.
The tools generated raw material — fast, cheap, endless raw material — and I spent the whole process doing the thing a director actually does: looking at what came back, deciding whether it did what the shot needed to do, and either pushing until it worked or building around the gap when it wouldn't.
Wrong movement. Wrong rotation. Wrong label. Wrong orientation. A shot that just never arrived. Every one of those was a judgment call, made by me, about what the commercial needed to actually say.
Sometimes that meant another generation. Sometimes it meant manually correcting the source imagery and trying again. Sometimes it meant changing tools. And sometimes it meant accepting that the shot I had in my head wasn't going to happen and solving the problem in the edit instead.
That's the part that doesn't show up in a slick case study. The finished commercial looks intentional because eventually it was — but it got there through twelve long, ugly fights with physics and label placement, not one clean pass.
If you're building anything with these tools right now: budget for the fighting. It's not a sign you're doing it wrong. It's the job.
(And yes — next time, I'm sticking to one image generator and one motion generator, so I'm only fighting one set of quirks instead of three.)
Comments