How I use Veo3 + Sora 2 to create Viral AI Videos (300M+ views)

0:00What if I can show you the exact workflow for how to come up with AI video ads that get hundreds of millions of views? Well, I can. Because today I brought on the viral AI ad man-man himself, PJ Ace. This guy is the number one guy when it comes to creating the most viral AI videos. and today he shows you his entire workflow, all his prompts, all the tools he uses, VO3, Sora 2, this app I never heard of called Rev, ChatGPT, his entire workflow, how he uses Figma and we just go through it and there was nothing that he held back. There are going to be a bunch of people make it to the end of this episode that learn how to create viral AI videos So I can't wait for you to enjoy this episode.

0:51People charge thousands of dollars for this sort of sauce, but on the Startup Ideas podcast, it's free for just a like and a comment. Enjoy.

1:07PJ, by the end of this episode, what are people going to learn? They are going to learn how to make a video that gets 230 million views, if you're lucky, but our entire end-to-end process that we've used to scale to a six-figure, seven-figure agency in like a few months. It's like been the most insane liftoff of the building of an AI native agency. And I'm excited to teach everyone exactly how we do it step-by-step. Okay. Keyword exactly, because we have a word on this podcast called sauce, and we don't like to gatekeep the sauce. So PJ, can you commit to spewing as much sauce, sharing as many prompts, sharing your screens as much as possible so that at the end of this, it increases the probability of success to actually create one of these AI videos that can go viral?

1:56I'm full open kimono. Don't point the cam down. Not wearing any pants right now. You guys are going to get full exposure with this. Okay. All right. Let's rock. Sweet. So Greg, you had me on the pod a month or two ago. I shared about how I made that viral video for Calci that got 50 million views. And, you know, today I want to dive deeper because we have better tools. Obviously, you know, the exponential growth of AI tools is like every other day we get, you know, a 2x improvement in quality. So it's pretty nuts. So I've got a lot more tools in the tool belt than I'm excited to roll through. We recently had an ad that did like for David Beckham that did like 230 million views.

2:33But the process for that was probably a little too complicated. So what we're going to do is do a little bit simpler one for one that a company called Origin Financial. This got 2 million views. So I'm really excited to basically dive into this ad. So first, we're going to watch the ad, and then I will take you guys step by step through it. Beautiful. Let's do it. All right. In the past, it was easy to get bad financial advice. It's a timeshare in beautiful Pompeii.

3:08Get your own cake! They said there was gold here yesterday. 20 % off on the first unsinkable ship.

3:22This is your college fund. Of course you're diversified. You have Enron and Blockbuster. I made all my money in meme stocks. Hey, get off my yacht! you've had money questions for years origin finally has answers two boat rides for the price of one

3:47score so good so my favorite part about it is just how extra it is yep yep and that's that so that is what we tell our clients from the start and so you know like at genre we we typically work with like larger larger companies they want to use ai video but they're not sure like how do we do this and not get the pitchforks drawn out because people are like you know you're cutting jobs or anything so we we always try and lead this is going to feel like a super bowl commercial in that it's going to be ridiculous it's going to be comedy first and there's going to be like mostly jokes and a little bit of brand at the end that connects it to you guys this is the only way we've found to like mitigate a lot of the kind of like ai disruptions changing everything customer pushback and so like you said if people are laughing kind of the whole time or they're entertained one they're more likely to watch the full ad which is what brands want and two they're less likely to bust out pitchforks because clearly this is like stupid it's ridiculous but it's kind of funny and you're like oh okay i get i get what they're trying to do so i think a lot of people have played with ai video but they haven't been able they want to do what you do but they don't know how right like it feels like they've scratched the surface, they've tried it, maybe it didn't work how they thought it would work.

5:04So what is the strategy for actually, like I've seen the result, how do I get to that result? Yeah. So we break it down into steps. Step one is scripting. Step two, well, let's just pretend you want to be a filmmaker and you want to work with clients. Everyone first always asks, how much should I charge, et cetera, et cetera. And the rates vary. At starting out, You maybe should charge nothing, build up a spec portfolio of like, you know, essentially targeting a niche of like, I love drink companies, beer companies, et cetera, perfume companies, clothing companies, et cetera. I always say like follow your interests to whatever videos that you like to make.

5:42You're naturally going to want to play with it and stick. You'll outlast your competition because it feels like work you're following. It doesn't feel like work. You're following your curiosity, et cetera, et cetera. So start in that domain and brands and then reach out to brands that are lookalikes, competitors, etc. And always kind of serve a niche initially. And then you can expand out from that. Like the first video I made, it was a pharmaceutical company video. And I had a ton of pharmaceutical companies reach out. But then I was able to kind of expand my portfolio with Calciad and then some other stuff to where we now can do a full suite of services.

6:13But anyway, you start off on the scripting phase with a given client. And usually we try and pitch three concepts. Now we work with pro writers, like I'm a good writer, but I'm not a great writer. I'm a good director and I'm a great director. So as you kind of expand over time, you do want to work with experts so that everyone's kind of focusing on their genius. And then eventually you kind of play the role of executive creative director or kind of producer that is orchestrating, you know, all the talent together. But I do think when you're starting out, you got to learn to wear all the hats and be kind of an army of one, which is where most filmmakers or creators kind of have that background of, I know how to dabble in enough of this.

6:47So anyway, step one is the script, you know, and so we kind of went through last time the script. Let me pull up the script for this one. So to give a brief overview of the entire process very fast, it starts off script. It goes on into ChatGPT to take the script, turn it into a shot list. It then goes on into creating a Figma board for laying out all your images and then generating all the images. And then you move on into a VO3 like animator to actually animate all the clips. And then you put it into an editor. So that is what we're going to be covering over the next 30, 45 minutes today. So let's jump in to the script.

7:29First, you know, you have to start with like the big idea for the spot. So the big idea for the spot is we like recognizable IP to start so that people have the familiar with the foreign. So for this, the opening shot we wanted was this kind of guy in Pompeii, which is, you know, iconic. And then the mountain's about to explode in the background. So that's kind of what we have in this opening shot. Actually, later on, we had the pyramid scheme, which was kind of a funny. He's like, eh, you like? And then the opening shot here is like an exploding volcano. So, you know, the idea is that like he's trying to sell a timeshare.

8:06And the kind of conceit of this whole thing is that in the past, it was easy to get bad financial advice. So it's like our writing team sat around and we're like, all right, what's the worst financial advice we give out throughout history? Let's try and make it in kind of a chronological order that leads up to today. So, you know, initially we thought, you know, OK, Pompeii would be funny if you're trying to sell a timeshare like on the foot of the dormant volcano. know, maybe something with Mary Antoinette where she's, you know, maybe she wants to buy more cake and there's like rioting soldiers outside.

8:40It'd be really funny if we could have like a Titanic moment. It'd be funny if we had like Beanie Babies is like the investment thesis for your college fund. You know, of course you've got your diversified, got Enron and Blockbuster. And then of course we wanted to end in some sort of meme coin, you know, kind of stab at all the the NFT bros. So, you know, always the, the writing just starts off like basically loose concepts with chat GPT. It's like, Hey, chat GPT, give me ideas for like iconic moments throughout history that incorporate like bad advice. And then give me some, like some potential lines for each of these.

9:15So then, you know, it'll spit out like, okay, here's like a bunch of different suggestions, bad advice. And chat GPT is not funny. 50 times out of, you know, kind of like whatever, Like 99 times out of 100, it's not funny, but one time it's good or it'll spark an idea that actually gets one of these great lines. And a lot of times it's not even like the line is funny. It's just the kind of comedic contrast where like the Beanie Babies or College Fund or the final bit is like two boat prices for the price of one. As the guys on the Titanic life rafts, it comes out as you iterate. And a lot of times you'll do half this commercial and then more ideas spring to mind of how you can dial it up, elevate it.

9:57So does that kind of make sense? Yeah. So what I'm learning here is there's three ways to get people to continue watching. One is using existing IP that people understand. People know Marie Antoinette. They know Pompeii. So thinking about and these are public domain IP, right? You want to use public domain IP. That's one. Two is juxtapositions. So how do you incorporate juxtapositions in a video? Because that's going to get people to share it and do something. You don't want them to lean back. You want them to lean forward, and that helps. And third is leaning into internet. What's trending maybe on the internet?

10:48obviously beanie babies aren't trending, but the meme coins are trending, right? So you're more likely, like the fact that you ended with that was a very smart strategic move because you shared it on X and a lot of people trade meme coins and talk about it on X. Yep, yep, exactly. And this feels like a little timeless. We wanted it to feel funny. I'll pull up real quick one of the other videos that were kind of similar in nature, but different. So this was this like Kalshi video and we don't have to watch it, but basically, you know, the conceit for this was like also historical moments. So we had like, will Jesus rise again?

11:29Peter, you know, the British are coming. Are they coming? It was basically odds throughout history of like different underdog moments. You know, like, will the Trojans accept their gift? Will David defeat Goliath? You know, Wright brothers here. And this one was not really funny as much as it was kind of like inspiring and it was big dramatic music and boom, it's your turn to defy the odds. So this is kind of a similar structure framework to tie something to a brand that feels historic and feels like, you know, universally relevant, if you will. Sounds good. Yeah. Yeah. So that's a similar kind of script and framework.

12:07worked. Um, so yeah, let's, let's, so let's just say you've got, you've got your script. The client is like, Hey, I like this big concept. So then you move on to the next phase, which is the exact scripting phase. And then we, we don't work with like scripting software. We just do everything in Google docs because clients can make notes on everything, et cetera, et cetera. Um, so yeah, once, once you've basically locked down your scripts, we need to move into the next phase, which is taking it all into chat GPT. So what chat GPT is going to be able to do is you can upload the script and you say, hey, give me a prompt for each of these as images.

12:42So we like to work with images on the basis that it's a lot easier and cheaper and faster to generate images for the commercial than it is to do everything text to video like you see here, where you're basically generating the entire commercial kind of blind and you're not sure if the client's going like it. So it's a, it's a huge benefit to be able to basically start to do everything shot by shot. And so that's why we basically take the script. We do scene one, scene two, scene three, scene four, scene five, all the way until the end. And then we start to fill it in with, you know, kind of each drafts of the shot.

13:21Now here you're typically what you're going to see is like all of the shots that led to this, which is what I'm going to show you in a second. So, you know, Mary Antoinette had like, she's got different, you know, kind of poses and stuff. And you would typically see that reflected in here. You've got lesser versions. But anyway, so we're going to use a platform called a Reeve and Rev. It's just app.rev.com. And basically it's pretty awesome because it gives you three different versions of whatever you're wanting to prompt. So we go into ChatGPT. I'll basically say, hey, ChachiPT, here is my script.

13:57I need you to basically turn this into scenes and a shot list for each scene. And I needed to give you shots. And I want the prompt to be structured. Let's just say I have like a master prompt kind of a thing that I use for all my things. These days, it's not that sacred anymore. The image models are great and they'll get you good stuff regardless. but you know it's like a nerdy and nervous Roman real estate agent strides you know arms out conversational gesture blah blah blah so it's going to describe a lot of motion but we're we're only doing images to start with so you know we kind of got something like this I lost the the real photos but as you can kind of see a lot of these images are quite similar to what we had in the final ad.

14:44So for all intents and purposes, these are the same prompts. So we'll kind of go through it. Now, the good news is once, let's just say we paste in a Pompeii thing here, or let's go Titanic. So we'll paste it in a Reeve. And then what we're going to find is Reeve actually gives you three variations of the images. And now thanks to Nano Banana-like technology, which this also works in Nano Banana, I just find Reeve's interface to be a bit better. and we also like the kind of photo realism we have on here you can even run it through other uh enhancer ai to make the skin more detailed and all that kind of stuff so anyway it's going to give us uh frames here now what it's going to say is it's going to suggest can we move the passenger uh closer to the camera can we add more dock workers can we show the passenger so it's like okay give me actually let's suggest let's click this so we actually click this photo and give me this as a closeup.

15:44And then, and then you can just hit enter and it'll actually give this image as a closeup because it's referencing the image. And then it's actually going to do three variations of it. So now the editor is, I'm sure as you remember, this is light years ahead of where VO3 text to video was back in the day. You didn't, you know, you to burn, you know,$4 a shot and you, you know, you just kind of pull the slot machine blind, but now you're really able to kind of go shot by shot and then start to build out. Like I said, each shot sequence of like, okay, I like this for the wide shot. I like this for the closeup.

16:20I like this for the peasants looking in through the door. Then I'm going to cut back to her. Is that, are you kind of tracking with me so far? Yeah. I mean, it makes so much sense by the way, to do script to images, then the video. In retrospect, we were crazy for going just straight. Straight into it. Just raw dogging text to video. That was crazy. The other thing I was thinking about is this seems better than Nano Banana, this Rev app. It's pretty good. There's a lot of realism and the structure feels right this chat interface uh you know sometimes it'll get like this guy looks like a fucking psychopath so i would uh re-roll this character but there's a lot of like this looks super realistic like this doesn't look ai generated at all um and and again to to you know do it it's like let's make this character uh more of a wide shot and include the captain yelling at him uh you know so it's just very iterative and for non-filmmakers like you really don't have to know a ton about angles and lighting and camera movements and all those things it just kind of builds it out with you conversationally which is kind of the future of of this um unless you go into image model world models i think we're going to see that start to infiltrate this stuff uh soon but anyway if if that kind of makes sense basically that's what we do is we basically go into copy code uh paste code copy code, paste code, copy code, paste code, you know, just back and forth until we have kind of all the shots here.

18:01Yeah, and I mean, you know, it really starts, you can kind of regenerate each of the images, but you can kind of see how you can build out multiple angles, coverage of a scene, consistent characters, et cetera, et cetera. You've got, you know, the guys smoking in the thing. And that's, you know, that's essentially the core of the image generations here. Cool. Okay. So we've got our images. Yep. So we've got, we've got all our images, you know, we're going to save them. I think we just hit the download button here, and then we're going to start to slowly put all the downloads into kind of our, our master boards here.

18:40Like I said, typically I'll have a bunch of like alt versions here because how, how we work with, we work with a dedicated writer. We work with a dedicated director, and then we work with AI cinematographers. And these AI cinematographers are the ones who We're going to take the script and the director's treatment, and they're basically going to generate a lot of these images and kind of fill out these boards with multiple options. You can see it here. This is for this David Beckham project we did with IM8, where the director will say, OK, I need this eyeball. They'll just upload a reference shot of an eyeball.

19:12But they're like, you know, we want it to be unique to the project. And boom, the AI editors churn out a bunch of examples or this opening shot. We need like a cool sci-fi setting, multiple angles of coverage. And then our AI cinematographers just do a bunch of shots, Serengeti, et cetera. So this is for that IM8. So as you can see, the more complicated the ad, the more, you know, like you can have hundreds of shots for just one shot on the shot list. And I do find that like when you're doing more stylized stuff, like this IM8 David Beckham ad that we did that got like 230 million views, you want to give a lot of reference images here for the tone the style so this is what this is kind of like if we're doing like the basic version this is like expert mode um you know you'll have like the line of dialogue here red or green it's a simple choice so it's like the first 10 seconds and then the director is coming in and he's saying wide shot of an abandoned facility we've got uh in this case we had the tennis star arena um and she was able to kind of we were able to deep fake shots of her um so you're just having like a lot of shots in this one setting and then our you know artists are doing these and then we're basically going to have the director come in and make selects of all these shots so that's what it looks like on more of a complicated you know shot so um that kind of brings us to we generated most of the shots on the shot list here we we laid them out to this here And then now we go on to VO3 animation as our next phase.

20:44Now, I do want to note that there's a lot of programs right now, and I'll list them off briefly. VO3 is probably the best model at the moment for making characters talk. It's just really realistic. The motion's great. Like, he's walking up here. Camera pans to that. So, you know, it's even worth kind of noting how we do this. So even though we have an original prompt for the Titanic, which is here, I don't know how to find that prompt. It's like this, you know, bursting dock, blah, blah, blah, snag tickets. So even though it's a text model, I'm still saying the dialogue. And then I'm just actually just copying and pasting that same prompt into VO3.

21:31And I'm uploading. So instead of text to video, we're going to do frames to video. and we're going to upload that shot that we just downloaded from Reeve and then NatX is our first frame and then that is going to be a similar prompt here just snagged first class tickets and it's going to be similar and ChatGPT can also help you it's like hey ChatGPT here's my image prompt now I need it to be like an animation prompt that you need to kind of tell me what's the camera movement here are we starting with him we're going to pan to the ship next and that'll help you really make these shots dynamic like you see here yeah and by the way to you pj because you're literally an expert you probably understand camera movements right to the average person listening to this including myself like i don't even i barely know what a camera movement means you know what i mean yeah dolly jib do a three-quarter turn yeah like honestly chat to bt gives you pretty great suggestions and you can also be stupid with it like move camera left to end on the ship.

22:33It doesn't matter. It'll do the same as some complicated director term. And so, yeah, I've found that VO3s is pretty good. Character performances are the best. However, if you're not having characters talk, that's when we'll move on into some of the other animations. To highlight, there's a platform called Kling with a K. Great 1080p, actually might be like 4k now another one called luma labs another one called um seed dream or seed dance which is from by dance which is from tiktok um great models another one called mini max honestly they're all kind of similar these days so that's why most people just default to vo3 because it will um do the best talking performances at the moment and it's google like their terms and service their indemnification like for working with clients a lot of times we'll just tell them Like, you know, if you just want us to use Google from start to finish, you're still going to have a great experience because Nano Bananas got you covered on the image generation.

23:34And then VO3 is got you covered on speaking and any animation. And Google's broader indemnification policies covers any commercial production. Like they're solid, great training data, ethically trained, et cetera. And even Gemini on scripts and stuff like that, right? Yeah, exactly. You can do this. And we're doing this for like an upcoming project for Google. like it's pretty seamless to go Gemini. Now, technically, I think you can even generate image, like text in Gemini, images in Gemini, and then even do video generations in Gemini. But if you're going to use the Google Suite, it's typically best use the Gemini app for that and then go to like labs.google for kind of images and videos and stuff.

24:17They also have another thing called AI Suite. It's Google's, they have great, always foundational technology, But the real struggle for them, I think, always is this application layer that sits on top is how do we make it seamless and cohesive? So they're still working on making this all like one big filmmaker suite. But in the meantime, Gemini app for or ChatGPT for images, I'm sorry, for the scripting phase, you can use for images either Nano, Banana and Google's AI suite. You can also use it on a platform called FreePick that has all of the image models. So let's actually just look through here.

24:53So in FreePick, you have, if you go to generate images, you can have the entire models from Google. You can use Nano Banana. You can use Flux, which is Black Forest Labs. You can use Seadream. You can use Ideagram. Like basically they have all the models in here. Now the downside to using an all-in-one platform for image generation and video generation is they're using API pricing. So it's just going to be a lot more expensive than if you were to go in. But there's a number, like for instance, image generation these days is practically free. So you get unlimited image generation on free pick for like, I don't know, 20 bucks a month, like it's nominal.

25:36And then I would just buy a subscription to Google flow. Cause basically for like, I think it's like 120 a month. You get, I think unlimited on fast mode, which is important to note that you shouldn't do quality mode. You should do fast mode just because it's like 80 % of the same quality. And I think it's free. And then you can do portrait. You can do landscape. And then the questions outputs. You could do four outputs if you want for that. And like I said before, just if you want to do frames to video, you can just add a starting frame here. And then you do your prompt, which makes the characters talk and et cetera like that.

26:13Beautiful. Yeah. Yeah. So that's kind of the process in a nutshell. And then obviously you take it into your editor. And as you saw, you know, essentially we're just putting, you know, generated clip, generated clip into the timeline kind of sequentially. And this ad was very simple to edit. You know, the image model or the video models like VO3 will add sound. They'll also add, you don't want it to add music, but, you know, it kind of like you just put a basic music track. It already has dialogue and sound on it. So the edits are like just extremely simple things. You really don't need to have a ton of editing experience now with everything so laid out for you.

26:51So you didn't – well, I guess, yeah, VO3 doesn't put on a music track, right? So you have to go and find – I guess either Storyblocks is the popular one. Yeah, most people use Epidemic Sound as well because for like$9 a month, you can get unlimited songs. And they've got a pretty big selection. The reason you wouldn't want VO3 or even Sora to like generate a music track is because it's going to generate a new music track for each clip because it's only giving you a piece of the pie. You want, you know, the whole thing to be no ingredients as far as music goes. But sound effects are great. And then how do you, this is probably a dumb question, but like what software are you using to actually put it all together?

27:34If you're wanting something free, you could just use CapCut is what a lot of people that are just want something simple. It's like an online editor made by TikTok. Quality. Most middle people like in the mid stage, I think DaVinci Resolve is actually free as well. And then most of the industry runs on Premiere, which is like 19 bucks a month, I want to say. I like Final Cut just because it's probably the easiest platform. It's like iMovie on steroids. But they all work. I mean, they're all great. And anything else we didn't cover on this whole process that you want to make mention? There's some interesting tricks.

28:14Like you can do these cool things where basically in VO3, instead of adding this as the starting image and then describing what Beanie Babies look like, you can actually include in the upload image, an image of the Beanie Babies that you want to uncover. And then like, you didn't even say a note that. So then when I uploaded, this is the starting image. And then I said, he uncovers this. Google actually outpaints this and he rolls it up to be the Beanie Babies that were there. So if you look here, this is kind of like a complicated thing, but it's just one of those details where he starts oh we kind of start with him midway here but basically the clip actually started with it fully closed but we had to give it a reference image for what beanie babies looked like and they had the big ty tags or whatever like that so just there's weird hacks like that that you can do where you can put picture in picture to describe what you're you're wanting to like pan the camera over to or if it was like a script like a like a if they didn't know what the Titanic looked like you could put a picture of the Titanic here and then you would say remove this image and then pan over to this ship and so then it would pan over to this ship and then it would look like the ship you did at the starting frame yeah this is crazy like even the beanie baby shot like if you if you I mean I just you know recorded this docuseries and I saw how expensive it is and how time consuming it is to shoot things like that shot would you can well you couldn't even do that shot right oh yeah i think you couldn't you know the the ty's aren't that big and it's so much more impactful with the ty's being big yep yep yeah this ended up actually being a different image but but it's kind of similar here and we had a hundred variations of this um because the issue is like you can't you want to generate this as kind of your end frame but you can't have it go in reverse to like pile in.