Coca-Cola's AI Holiday Ad Used 70,000 Video Clips

Coca-Cola's new AI-generated holiday ad has sparked both praise and controversy. In this episode, Joey and Addy break down the technology behind the commercial, analyzing how Secret Level combined tools like Veo 3, Comfy, and Sora to achieve CG-quality animation with a small crew instead of the large team usually required for traditional production
--

The views and opinions expressed in this podcast are the personal views of the hosts and do not necessarily reflect the views or positions of their respective employers or organizations. This show is independently produced by VP Land without the use of any outside company resources, confidential information, or affiliations.

Transcript

Welcome back to Denoised. Let's talk about the Coca-Cola ad. This is the third year I believe that they've used AI in their holiday ads. Ooh, Santa's super plasticy man. Comfy Runway. Veo 3. Kling. Sora. 

Sora was used to turn this into a meme.

All right. Before we start Addy, I saw, I haven't been to the theaters in a little bit and I went back to AMC last night and um, there is a bit of a travesty happening. I dunno if you're aware of this, but they cut down the Nicole Kidman spot, the Nicole Kidman. Movies are magic is now, I don't even know, it's like 10 seconds.

I think it was maybe what, 30 seconds before. But it was like, it played and it was this cut down version and I was like, what happened? However, it was long on, long enough for people. We have a whole episode about what happened. It was on long enough for people in the theater to still cheer and clap because AMC Century City is a quality theater with a quality audience, as we've covered before.

Some would say the best kind of audience, Hollywood insiders. Um, 

some of the audience did walk outta this movie though, so. Maybe not all of them are the best audience. 

I'm, I'm guessing just the rights to use that probably expired. You know, I think, I'm guessing it's 

because they wanted to squeeze more ads or more trailers in and, uh, they wanted to save time, the 20 seconds.

Oh, so it's a purely, also, they did not cut 

down the Coca-Cola ad that. Everyone sort of loves to hate that sort of riffs on the Fast and the Furious and other movies that I can't remember right now, which plays immediately. Cinema 

Going. Cinema Making as a religion is kind of getting hit by economics.

Yeah, but I mean, Nicole Kidman, you 

can't touch her. That's, that's sacrosanct. 

Oh boy. 

All right. Speaking of Coca-Cola. There is another Coca-Cola ad that is, uh, you know, causing some more online controversy as usual. Uh, it's another AI Coca-Cola ad for the holidays. This is the third year I believe that they've used AI in their holiday ads.

So yeah, let's run. I think there are two different versions. 'cause every time people link to stuff, and when I checked out Coca-Cola's YouTube channel, there are two different versions, but. They're both made, I believe, by, uh, secret Studio, uh, Jason ADA's company. 

Yeah. I actually have heard a lot about it over the last week.

As you know, you and I are both in New York and I was deep in the agency world over there, and everybody was talking about this commercial, so I saw bits and pieces of, but I would love to do a live reaction 

with you right 

now. Yeah. Let's play 'em. And I'm 

sorry, with Secret Level, Jason's at his company, not secret studio, Secret Level.

Here. Let's play. The holidays are coming fantastical. Version.

So animals reacting to Coca-Cola trucks driving. Mm-hmm. I'd say the details of the animals look. Was that Peter Rabbit? Really good.

Yeah. It 

has a very distinct Coke style to it. Yeah, I mean the qua, like it looks like a CG animation. The quality has definitely leveled up a lot. 

Yep. You could see the AI sheen in some of the shots. 

Yeah. Some of 'em look really sharp. That wide was a little, 

the hubcaps had a Coca-Cola detail. That's nice.

Ooh. Santa's super plasticky, man. 

But I feel like that's, that to me, that shot is the shot that make, gives us the reason of like why you would go in it. Why you would go ai? Because like everything else, it's like, okay, why not just do it cg, this shot is the reason I think like the AI makes sense. 'cause this is like a throwback to the classic like 1930s Norman Rockwell.

Yeah. Hand illustrated. Yeah. And they brought him to life, which is the the big thing that AI's cool at. I think this was probably the easiest shot out of all the other shots because it's just an image to video on a very basic motion. Yeah. I mean, I'm sure that it composite a lot of the stuff or not have been at a lot of the stuff to make the frame and then.

Uh, image to video. Yeah. Yeah. You wanna see the other spot? 

Yeah. I mean, before we get to that, yeah. The one thing that is notoriously difficult with ai, you and I both know this is text and specifically very stylized. Uh, cursive text that the Coca-Cola logo is. Yeah. Even there I would say that it's, uh, it's a little too thin.

So maybe like the aspect ratio seems too tall here. 

Like 

they Oh, like the perspective? Yeah. Even they didn't get it 

right. Like the dimension, not the dimension. The, um, like it's a little distorted. The aspect ratio is what you were. 

Yeah. Right. Like the, the, the C should actually be touching his finger on the left side of the frame and the A should be touching his thumb because the aspect ratio just has to be a little bit wider.

Yeah. You can go to an actual Coca-Cola logo like. Yeah, and if you look the, at those two side by side, yeah, so the bottle one seems a little bit thin. I know it's a super small detail, but people like Coke actually obsess over these details. So you can pause on any frame where the Coca-Cola logo is shown.

That feels like it was just plastered on, right? Like after the fact. Um, it, it just does, like the white color, the color temperature of the font doesn't match the color temperature of the bulbs or anything else. So that feels after the fact there.

Well, okay. I mean, I, I'm thinking first off, Coca-Cola signed off on this, so whatever minor discrepancies with the logo, I guess they're cool with. Look, I mean, overall it would, you 

wouldn't get that 

level of error though. No. You'd have the control, but I mean, that's still the. Debate. I mean, not the debate.

That's just still the battle with AI of just like the battle of speed versus control. I mean, I'm thinking for something that is so, has so many thin lines and is a pretty detailed logo, it held up really well for, for with ai, even given the current state of AI and going 

Yeah. That, that I'm, I'm curious on what the AI workflow was at the Secret Level because.

I know for a fact if you just throw the Coca-Cola logo into a video generation model it, it's gonna come out. It's gonna work. It chopped up. 

Yeah. Well that's why. And so, I mean, there is a behind it. So yeah, let's talk about that. 'cause they have a couple shots about with the logo and that gives a hint of the workflow 

and um, uh, I was just joking while the video was playing, but there was a shot of a rabbit looking at the truck driving by, and that rabbit looked exactly like Peter Rabbit.

The latest version of Peter Rabbit movie with, uh, beanie and the the ears kind of pulled back and I'm wondering if it's supposed to be Peter Rabbit and they had the IP license for that, or it was just. Video models generating stuff that they trained on? 

I, yeah, I mean, I did see some other comments online how some of the other characters looked like some characters from Zootopia.

But uh, if you look at the behind the scenes video, I mean, from what they're saying, everything was started from like character sketches and different, like, character sheets turned into, first frames turned into video from the behind the scenes video. It wasn't like, oh, let's just get a shot of the. Rabbit of like a shot of the rabbit generated and then it happens to look like Peter Cottontail 

Cottontail.

I don't know. That's right. 

Was that, oh wait, what were you saying? It's Pure Cottontail, right? 

Peter Rabbit. Peter Rabbit is the cartoon and the ip. Yeah. 

Okay. 

Okay. Here's the behind the scenes video. Given AI company A BTS 

task in 2025, Coca-Cola and Secret Level partnered again to reimagine The holidays are coming commercial.

This time expanding into a brand new 

story world with animated? Yeah. I mean, those shots are sharp man. Like they use some good caravan every year since 1995. 

The iconic campaign has signaled the official start of the holiday. 

Yeah, that's obviously composited our re Yeah, but that was for the behind the scenes thing.

Yeah. I mean, lemme jump ahead. It's, it works. It really works. Oh, go back to that wipe. 

Yeah. 

What, what are they wiping? Timeless icons and cutting 

edge technology. That's Photoshop. Okay. It 

looks like an image. Yeah, image to video. So look, it has these 

like things of some 3D models. I'm trying to wonder where the 3D 

models come into play.

Um, yeah, like that, that bottle's way too small, unless it's the, the mini 500, uh, 250 milliliter bottle. They've got this sort of, uh, this. 

Depth extraction thing. Um, yeah, I'm trying to recognize what software this is, and then here, this is a magnifi, upscale, nice. But here, look, they have character sheets, sketches, you know, they're saying everything started from sketches, started from artists drawing the sketches.

And, and sure, I could believe that turning into the animation, a lot of, in painting and the shots to tweak the first frames. It, 

it's, 

this is a complex challenge because, uh, here, and people are kind of losing their mind over this. Um, so look, they're compositing the logos. It still looks like they're doing the logos on the image, but they're compositing the logos into the image.

And that's Veo 3. 

This, uh, that's what I would do. 

Yeah. I 

wouldn't even try to generate the logos. I would just add it on after the fact, which is exactly, well, no, I'm 

impressed that they're adding the logos onto the image and then running image two, video in Veo 3 and that it is holding up. Oh really?

See, look, I mean that output was Veo 3 

Oh, there, there's that rabbit character. 

Yeah. See this out? I guess it's not Eater Rabbit. This is Veo 3 and that output and the logo. Looks pretty good. I mean, look, we just did a commercial where most of the shots were AI generated and it involved like children book covers and if the text is really big, it would kind of hold up, you know, in Veo 3.

But anything smaller like author names and stuff just turned garbled. But for the entire book covers, we would just re composite the actual image on top of the final video, which I'm wondering if you did that with some of the logos you, it's got look, I mean, Texas has gotten better, but it's still not great with.

Uh, yeah, so there's your rabbit character.

Yeah, it looks like a lot of very detailed, I mean, it's kind of interesting to me 'cause they're doing all this, like they're showing shots of like very detailed, like modifications on the frames. It's, it's, this stuff's gonna get shrunk to HD or lower in the video. And I'm just wondering, like, and then output at 720, maybe 1080, obviously you're gonna upr it.

I'm just wondering, does all of this detail painting really matter? Because it's gonna get, it does. It's gonna get. Compressed and respit out as a video that, that is way low fidelity. 

Yeah, I, I think it absolutely matters because there's just such a high number of furry creatures in there. And fur is notoriously difficult in CG as well as in ai.

Um, you know, you're talking about aliasing and moray and so yeah, I think you have to hand paint as much as you need to do. Yeah. Alright. Interesting. 

I think, um, I mean, the other thing that works with a fur and with these shots and from my experience with AI gen, how close 

were Zootopia? Um, characters for sure.

Um, so if you go to the shot after this right there, the koalas, I guess, uh, they, they would be, um, koalas. Yeah. Yeah. That looks just like 

the Zootopia Koalas. I, the thing that helps with the fur is they're all closeups or medium shots. Yeah. You know, the closer you get with the detail, the better the AI is and the more singular.

I mean they even have a 

anamorphic Boca on that Quala shot. I would with this shot. It's 

not consistent with the rest of the shots 'cause the rest of the shots seems spherical. I'm impressed with this shot 'cause it's a pretty wide shot with a lot of leaf plant, palm tree details, lot of detail there and the, the details hold up.

Oh, look at that. Yeah. Here you could see a bit and this shot of, uh, I don't even know what this like, kind of looks like Paris kind of looks like some canal shot. Um, yeah, I mean, look a little bit more of the world it sounds like. 

I'm being super picky here, but honestly, this is as clunky and low quality as it'll ever get, and it'll only get better from here.

Yeah. So if you imagine commercials of a future, like this is the prototype that they're gonna build on, and, um, I, I think you're just gonna see less and less. Resources, less and less touchups and faster and faster outputs, honestly, 

right? I mean, the flip of this is like, okay, cool. If we can make a high quality ad with fewer people less time or budget, then why is it just making one ad?

It's like, well, now you have the chance to like make multiple variations and multiple holiday ads that kind of live in this universe. Maybe that's why they made two. I mean, for the same budget. I, I'm, I'm just guessing, I don't know what the full budget was, but, um, you know, I think that's what is the promise of this, of being able to do this kind of stuff, you know, faster and cheaper, make more stuff.

Okay. 

I, I, I love, I love that thought. My thought in addition to that is. Is Disney looking at this for Zootopia 3, 

I mean, I think every animation studio is looking at this at, at some form in their pipeline. Like, you know, we talked about this before. You know, I mean, I don't, I think people get this misconception of just like, yeah, I'll just type on a text this video and make a Zootopia.

And it's like, no, this is the stuff to help you just save the in-betweener shots or compositing shots or other stuff in between. You still see the amount of, um, work that goes into it. 

Yeah, and, and like the whole look and feel of the Coke commercials of the past. I mean, all of the, like the ice and the nighttime driving and the nighttime shots with like a, like a dark blue, you know, indigo night sky with a bunch of stars.

Like all of those elements that make a Coke commercial, what it is, is retained here. Very much so. It just is so much more complex than I thought it would be for an AI commercial. 

Like complex 

in the work that 

went into 

making it as well as like the stuff that is in it, like the number of characters, all of the different shots, the different environments like this.

This is a lot, even for a traditional production, 

my favorite reaction that I've seen online, I've obviously, there's a lot of, you know, pushback and just like, ah, AI is like stealing jobs or yada yada. Um, but this one it has, oh, now it has over 164,000 likes. But, um, yeah. Matt Lieb, basically, uh, Coca-Cola used AI to make their holiday ad, and when people got mad, they created a fake behind the scenes video, uh, with a screenshot of the Coca-Cola logo being, um, composited onto the frame in Photoshop.

Uh, and people say, say that, that's not how this works. As if, I don't know that's cheating with ai. I don't, I still don't fully know what that commentary means. But yeah, I think, I think, 

uh, the mainstream folks like. Middle of the road folks who don't mess with AI as much as we do, don't have idea how easy it is to combine AI workflows with traditional workflows like the, they compliment each other, if you think about it.

Yeah. Or I don't know, like even there that there's some, a purity of like being AI only workflow and it's like what it, whatever tool gets the job done. Uh, it gets, it gets you to the end result that Right. In the time and budget that you have. But yeah, as we've known, we talked about extensively before, there's really, um, no way around getting good detailed outputs with things that you need to honor, like logos and text, uh, without compositing it manually.

I'm impressed if they were able to composite it in the fir in the frame and get final output at a Veo or whatever model that's, that's impressive. And not have it warp or completely change? 

Very much so. Uh, look, uh, like when we were covering the Tilly Norwood stuff, I think I was, I definitely sided with the public on why I didn't think that that was the right move at the time.

This one, I, I feel like it's pretty spot on and. It's, uh, it's timely, it's relevant, and it's using the technology in the right way. Having said that, uh, there is a comment from, uh, on x from Vashi Nedomansky, and, uh, it's the second link, Joey. So it's got all of the, the, like the stats on how long this took and how many people it took, you know, something like 50 AI specialists, 46 hours per person, and a hundred people involved in the project.

Where do they get this data from? Is this accurate? 

Vashi is a Vashi a professional editor. Uh, yeah, he was trying to compare what this would've cost if you did this with traditional VFX pipelines. Some of these numbers are accurate, but yeah, I think that's sort of jumping at yes, this isn't a good order.

Um, 'cause the thing I wanted to get, ATTN this number is true. They generated 70,000 clips with Comfy for the spot. Sure. That part, I really don't understand how or why you would need to generate 70,000 actual clips. I mean, maybe first frame images, that's still a lot, but I would understand that more generating full video clips, 70,000 of them.

I'm trying to, I would love to know, um, why, 

wow, somebody was really happy with the API money. I don't 

when you get unlimited, uh, access to Veo 3. Yeah, I'm still very confused about that. 

So let, it seemed like there was maybe about 20 or so shots in there. Right. And if it's like. I don't know. 50 50 iterations per shot.

You're still looking at like a thousand generations tops. 

Yeah. And there's no way you would need to do, and that's generous. 50 iterations. Yeah, if you're, I mean, the stuff's gotten so good. If you're prompting and giving it some direction, maybe five to 10 generations a shot. Totally. But you should get what you want.

Um, yeah, that's a very, how you get 

to 

the 70,000 number. 

Yeah. 

I don't know. I'm so confused. Yeah. That this, I saw this number and it was in, one of, it was in one of the Hollywood reporter, one of the articles, and Comfy tweeted it out too. I mean, so they, they backed it up. Um, I don't know. I'm very confused about that.

Yeah. The Vashi tweet, uh, the 70,000 is, you know, that's accurate. Yeah. These numbers. It's, uh, they did say that they used 20 people on the crew, which was smaller team than the spot last year. Yeah. I think he was trying to price out how this would've compared, which, no, I think he, he posed the question. He didn't actually, there's no number behind that, so I, we don't know.

We don't know what the budget is for this either. They said it was less than if they did it with traditional methods, but they didn't actually give a number. 

Yeah, I would imagine it's still, uh, significantly cheaper and smaller in team size than, uh, traditional computer graphics. I mean, just building those different environments.

We saw the palm tree environment, the, uh, Chinese Panda Bear environment. We saw the, the. Truck driving through the snow, like I counted maybe as many environments as there are number of shots. So let's say 20 environments, each one of those would require at least five to 10 artists per environment. Mm-hmm.

To build these out CG traditionally. Yeah. So you're already looking at, you know, maybe, I don't know, a hundred environment artists and then you need character artists, animators, riggers. Uh, because there's just so many furry character. You need people to build a fur. Mm-hmm. You need people to get reference from a real performance and then key frame, animate all of that.

So, uh, maybe another 50 or so people just for character animation altogether. And then you obviously need editors touch up and their traditional stuff. Right. Post-production stuff, you're looking at a team of about 200 easy. For this, if not bigger. 

Yeah. And whatever. Budgetwise, take that, whatever. The 20 people here multiplied by 10.

Assuming similar day rates or week rates. Yeah, roughly four times budget. It depends on four to 10 times as much. 

Easily. It depends on where on the planet you were making this. Right. So obviously budgets are gonna be a lot smaller if it's in the Philippines or India or someplace like that. But if it's like US based talent, I mean you're talking about five to $10 million.

All in. Yeah, with all of the approvals with Coke and you know, the stuff that they don't show they don't use and so on. 

Yeah. The other, um, 69,000 video clips. The, the bottom of the iceberg. Yeah. On the other behind the scenes spot that I skipped through that I said sound like a notebook, LM episode, uh, they did show a couple more behind the scenes shots.

Sheet of different tools and it's sort of basically the usual suspects kind of all thrown at it. Uh, Comfy Runway, Veo 3, Kling, Higgsfield, Krea, Sora. I'm curious where, what Sora was used and where Sora was used for this PAC to, I haven't heard that one. Sora was 

used to turn this into a meme. I'm just kidding.

Uh, this is like such a cool image to see here. It's like the, the United Nations of AI models. Um, I don't know what PTO is, but I 

dunno what PTO is. Yeah. Yeah. And yeah, there's one, I had another shot of, uh, using the Veo 3 text, uh, uh, markup to steer the shots and some Comfy shots here. Comfy, uh, node.

Nice. All roads lead to nodes. Yes. At least this is where they mentioned 70,000. Clips. 

Oh, got it. Okay. Yeah, I don't think people saw the other 60,000 clips. They were just generated just because, 

I'm just curious why. Yeah. I don't know. Yeah. Unless you, they have some agentic workflow that was just like spit out stuff and we'll just select what we like the best.

Yeah. So I don't know. What 

are your thoughts? What are your big 

takeaways? Look, I mean, the fur, the details look great. I think Frame.io, it still feels like. It, it could have leaned more into what AI was good at, which was like bringing 1930 Santa Claus to life. The other thing is there's no speaking characters.

That's still a weakness. We did some on the spot and it's, yeah, it's, it, it's still a little bit uncanny valley and so there's still, you know, avoiding anything. Speaking. It's all voiceover or music. Also speaking to music, they had an original jingle, which I'd be curious how that original jingle was created.

Ai. I, I doubt it. I bet it was probably commissioned. I don't know. I'd be curious. Yeah, I mean, look, I think that technically, um, the quality, you know, obviously compared to last year is way better. But just overall the spot, I feel like I would've wanted to see some more of what AI's good at. Right now. It feels like this was a good CG spot that they're able to pull off with AI tools.

Yeah. Yeah. No, this is, uh, what do you think? I think this is the way things are going. This is exactly the zeitgeist and the, the status quo as it is in, you know, end of 2025. Yeah. Coke is big enough of a company, uh, to obviously put resources in this and the trust in AI to do this. But also Coke is a very, they're very protective over their brand.

So for them to kind of lean in and just go all in on this, shows the direction in which everybody else is going. If you look at a brand like Apple, how protected they are over what Apple products look like, what Apple commercials look like, you know, they're next. Like, if Coke can pull this off, I think the, it's all, it's all good from everybody down.

Down that ladder. 

Yeah. And yeah, I wanted to call out a Coca-Cola. Oh, I didn't even realize this. So, uh. Uh, Pratik, Pratik Thakar, head of generative AI at Coca-Cola. I didn't even wanna say, had a gen dedicated AI role at Coca-Cola. He said, we need to keep moving forward and pushing the envelope, the gen out of the bottle.

And you're not going to put it back in about Yeah. Using AI for the ad 

for, for better or worse. That that is the truth. I know, I know what the implication of, of that. Like there's a lot of people out there who are gonna be left behind, but then there's also a lot of opportunity out there for you to.

Create a new path in your career. 

I mean, that's also what I'm saying, where it's like, you know, you're asking me about this specific spot, and it's like, yeah, cool. Okay. But they're able to pull off this spot as we, um, guesstimated that, you know, is 10 to four times less expensive than doing this.

Traditionally, Coca-Cola has the money. It's like, okay, now you can make more of these. It's like, you could make this animal. This quote, Zootopia, Coca-Cola universe, and have like stories about this for the entire holiday season. It doesn't have to just be your one hero advert, right? That you run repeatedly.

So like I would wanna see more of that. It's like, okay, cool. We could say we could do this and save money and. Let's make more of it. Yeah, hopefully. 

Yeah, let's get, I mean, again, the toothpaste is out of the tube or the genie out of the bottle, however you wanna phrase it. Like, Pandora's box is opened, Pandora's Box is opened.

We're not, we're not gonna go back to a hundred percent computer graphics ever again because some tasks are just inherently easier and better this way. 

Anything else on the Coca-Cola story? 

No, that's it. 

All right, we moving on the, the Coca-Cola cap of the glass bottle has been uncapped. It cannot be 

recapped.

I think I have some Coke in the fridge. I'm gonna have some after this episode. All this Coke talk got me thirsty. 

We're not sponsored by Coca-Cola, 

but 

open to it.

All right, thanks everyone. We'll catch you in the next episode.