On the budget question, the short answer is: yes. 
But I have a slightly weird perspective on this because I sit in the middle of all of it. Our clients have budgets and problems. Our agency writes scripts to solve those problems. Production companies figure out how to capture what we need, and then I work with production, our creatives and post to figure out how the hell we’re actually going to make the thing.
So I don’t really separate “production savings” from “post savings” as cleanly as maybe I once would have.
Shooting stuff is expensive. But shooting stuff is also important, and I don’t think the answer is simply “don’t shoot it, use AI.”
What having these new capabilities on the post side does is give you some pretty ridiculous superpowers when you’re making those decisions. You can do more with less, obviously, but more importantly you start to understand what you actually need to shoot, what you don’t need to shoot, and where it’s worth spending the money.
I’ve been saying this for years, even before AI: I think people who really understand VFX are uniquely positioned to help agencies and clients make those decisions. Now you add generative AI to that toolbox and the range of problems you can solve gets much, much bigger.
Basically: know when to shoot it, know when to VFX it, know when to AI it, and know how to AI/VFX your way to happiness. 
That’s much more interesting to me than “AI makes this 30% cheaper” or whatever.
On working graded vs. log:
Almost all of the generative work on this project was done on top of graded footage.
And yes, I know. Math is math. In a traditional VFX pipeline, working pre-grade is the “correct” answer. 2+2=4, and once I’m working on top of a grade, sometimes 2+2 does not equal 4 anymore.
But there’s a practical reason for it.
We’re a tiny team of X that 8x-10x freelancers when we need to. We’re basically a group of expert generalists, and color isn’t our core business. We rely on best-in-class color partners to do that work.
The amount of back-and-forth required to run all of this generative work pre-grade would absolutely destroy the schedule and budget.
And there’s another important difference: these AI models are effectively very comfortable with graded imagery. That’s the world they’re used to seeing. Working with the grade baked in is easier, faster, cheaper and, most importantly, produces results that are good enough to make the overall thing possible.
We still have all the camera originals, obviously, and we’ll ACES our way to happiness wherever it makes sense. But the generative work itself was overwhelmingly done post-grade.
Is it perfect? Nope.
Sometimes the color shifts a little. Sometimes we live with an imperfection. Sometimes a creative makes a decision with the colorist today and tomorrow I have to move something 3% because it makes the generative shot work.
But that’s the trade.
If I can go back to the creatives and say, “Okay, occasionally we’re going to have to accept some little imperfections or make some compromises — but in exchange we can make this entire long-form thing that otherwise simply wouldn’t have been possible,” that’s a pretty easy conversation.
Before/afters:
I would love to show some. I just have to be really careful about what I share. I’m going to see what I can put together that doesn’t step on any landmines.
And Nick — the traditional roto shot:
It was the touchdown catch.
Funny enough, that also turned into probably the hardest shot in the entire piece.
We used video-to-video to turn a completely different part of the football field into the end zone, and the models just hated the action.
Apparently generative AI finds the concept of a football player running without a football deeply offensive. It desperately wanted to put a football in his hands whether we asked for one or not.
We spent more time fighting that shot than probably anything else.
And it’s also one of the shots where, knowing what I know now, I would probably approach it completely differently. I’d likely do it with more traditional matchmove/projection techniques, maybe some light CG, and use the generative tools much more selectively.
Which is actually one of my favorite lessons from this whole project:
Just because you can solve something with generative AI doesn’t mean that’s the best way to solve it.
Sometimes the 25-year-old VFX technique is still the right hammer.
