How to Automate Viral TikTok Videos

Transcript

00:00

Why? Why?

Why? Why is that the B-roll it chose?

Have you been seeing these like science explainer videos pop up in your feed? You know, the ones where they're like,

00:21

they're like one step away from a Coco Melon song. The most popular I've seen is from a Tik Tok account called Learning with Lyrics.

And these guys are kind of killing it right now. These things are super viral, like autotuned songs with lessons that feel straight out of a high school textbook.

And for

00:38

some people, this has clearly become a pretty lucrative side hustle. Personally, I've never really wondered how an air conditioner works.

But damn it, I just spent 3 minutes learning all the ins and outs of climate control thanks to these stupid addictive videos. But by now, you've probably guessed it.

00:53

I like side hustles and I like money. So, I want to make one of these accounts for myself and start raking it in.

I have a pretty solid idea of how these creators are putting these together, too. Unfortunately, I just haven't had the time to create them.

So, effective immediately. No, I'm just screwing with

01:08

you. But, I did think it would be a fun idea to try to recreate these things, show you exactly how it's done, and possibly even put the whole thing on autopilot.

So, let's go ahead and dive in. Here's what we need to do.

First, we need to research the topics. Then, we need to write the songs.

Then we need to actually create the songs so that

01:24

they're poppy and catchy. Then we need to figure out the visuals for it.

And finally, we need to pull it all together into a single video. And then it's just a matter of publishing it to our social media accounts.

Now, there's likely going to be some manual intervention required, but let's create one video first using all of the best tools. And

01:40

then once we have one video, we'll try to pull it all together into a nice mostly automated system. So, let's go ahead and get started with research.

Here's

01:55

a prompt that I have found works for actually doing the research and generating the lyrics all in one step. Explain to me exactly how a car engine works.

Keep it simple but thorough so anyone can understand. Give it to me in the form of a poem.

Avoid using

02:12

metaphors. I feel like avoid using metaphors is important because when you tell it to write a poem, it's going to try to rhyme as much as possible and will use metaphors and sort of shift the meaning of things to get it to rhyme.

In fact, I'm even going to add it does not

02:28

have to rhyme. Already done.

How a car engine works. Air enters through the intake valve.

Fuel is sprayed into the cylinder, etc., etc. I'm not going to read the whole thing, but this is a simplified version of how an engine works.

02:46

Now, in my opinion, the best music generator out there right now is most likely the videos that are going viral probably use to generate them. So, for this specific song, I will use Sunno.

However, we are going to run into an issue down the line using Suno cuz Suno

03:03

doesn't have an API. So, it doesn't connect to all the various automation tools where you can just tell SNO to generate a song within like an automated workflow, but Sunno generates my favorite AI music right now.

So, we're going to stick with that for this step and we'll figure out how to solve that

03:20

other problem later. Once I'm in, I enter some prompts, play with the settings a bit, which can be fine-tuned to your personal preferences, and then hit create.

And this is what it spit out. Fuel is sprayed.

[Music]

03:36

Not bad. But I'm going to go ahead and generate another variation.

I mean, it's actually pretty catchy, but it's not quite the style I'm hoping for.

03:53

Yeah, I don't like that one either. So, I did a few rerolls until I finally got one that I liked.

And I really like the way this one sounds right here. enters through the intake valve.

Fuel is sprayed into the cylinder. A piston moves down.

We're drawing the mixture

04:10

in. Like that's really kind of what I'm going for.

Very catchy. It's the right voice.

It's slow enough that you can make out what she's saying. And so one key inside of Suno is once you find a style that works and you want to make sure it sounds like that same musician

04:26

in the future. Well, if I come over to these three dots on the song that I do like here, I can click on create and make persona.

And then we'll name it Tik Tok education songs. Now, if I want to create another song and I want it to sound very similar in style, I set some

04:41

new lyrics. I click on persona.

I select the persona we just created here. So now every time I go and create a chat GPT educational poem and pull it into the lyrics, it'll sound like the same band or same artist making that song.

But now

04:58

the next step is quite honestly the most difficult step. Both if you're doing it manually and also probably the most difficult step to automate later.

But we need to get visuals for this whole thing. I've always thought it would be kind of cool to start some local business like a bar or something.

But

05:13

before any business can actually take off, you need branding that stands out. So today, I'm going to design a whole brand for a modern cocktail bar using Recraft, the sponsor of today's video.

ReCraft is an AI image generation and editing tool with over 5 million users worldwide, trusted by brands like

05:29

Netflix, Oglev, and HubSpot. And they have this new feature called chat mode.

It basically splits your screen in two. So on the left, you can chat with the AI, and on the right, you've got a canvas where your ideas in each iteration come to life.

and it'll actually remember each iteration as you

05:44

work. So, watch this.

I just type logo for a futuristic cocktail bar called future proof. Sleek typography, neon glow, minimal modern style.

Instantly, I've got professional logo concepts. Then I can refine the design using the chat mode feature.

Make it darker, add

06:00

more modern accents, try a teal and magenta color palette. And you can see that it remembers the first iteration while updating it with what I asked.

And Recraft isn't just for logos. I can drop the design it made for me into chat mode and ask create posters, coasters, and

06:15

social visuals in the same style. In a few seconds, I've got a full brand kit that I can make merch with and crossost to all my social media pages.

So, if you want to try this out before anyone else, join Recraft's new chat mode beta weight list. The links in the description.

Now, let's get back to the episode.

06:32

[Music] One way we can try to do this is to use Nvidia. Nvidia has been a sponsor on the channel in the past.

They didn't sponsor this video, but they have a handful of features. Now, the feature I've talked about in my videos in the past has

06:48

always been their generative AI videos where it tries to generate the entire video. So, I'm actually going to use the old features that Nvidia was designed for to see if we can get it to find stock footage for the script that we have.

All right, so jumping into Nvidia.

07:04

First, I choose which social platform I need assets for. And then I choose what type of assets I need, the platform, the type of media, and any additional guidelines I want to give it.

In this case, I chose Tik Tok stock media, and then asked it to keep the total video length to 90 seconds. And a couple of

07:20

prompts later, some of it's usable. Air enters through the intake valve.

Fuel is sprayed into the cylinder. A piston moves downward, drawing the mixture in.

The intake valve closes. The piston rises, compressing air and fuel tightly.

I have no idea what was going on with some of the music in there, but there

07:37

were some visuals that are usable. I like this visual right here where it shows the piston go up and then as it comes down, you know, it gets its fuel and oxygen and creates an explosion.

Some of this footage has absolutely nothing to do with what's being said in the script. For example, this says the

07:54

crankshaft spins right here, but it's literally a still image of a few cogs, but the cogs aren't spinning. And this I don't believe is actually a crankshaft.

It almost actually I don't know if that's the crankshaft or not. I I still don't know engines very well.

08:12

[Music] Now, the absolute best video generator of the moment is Sora 2. So, let's take some of this text and plug it into Sora 2 and see what it generates for us.

And then under describe your video, I'm going to paste in those three lines.

08:29

Make sure we're set on portrait. So, it's generating in 916.

And let's just generate based on those three lines and see if it can come up with a decent visual for us. Now, I'm not totally getting what I want.

This one is, you know, almost correct. I don't think

08:45

these two things overlap each other. And if this is supposed to be the intake valve, it's supposed to be, you know, sealing the gap up here.

Not quite right. The one on the left is a little more zoomed in.

I don't know why there's air or whatever going below the piston here. That doesn't make a lot of sense.

09:01

But the closeup kind of is cool with it shooting fuel in, but then it it just doesn't make a lot of logical sense. [Music] The next option would be to find images

09:16

of what I actually do want. So, I did a quick Google search and I found these screenshots here of like the intake, the compression, fuel injection, and exhaust valve.

And so, maybe I can get these images animated for me. So, I'm going to

09:33

use Leonardo to generate with V3, but go ahead and use whatever platform you prefer for generating with VO. I just find Leonardo to be the least expensive.

And make sure you do have V3 selected here. We'll select a start frame.

pull in the image that we just found. And

09:50

just for fun, let's have it generate with the Cling model as well. I'll use Cling 2.5 Turbo and we'll generate with that model.

See which one does it better. And well, here's the two sets of visuals that I got.

Neither of them are

10:05

very good. V3 kind of added some funky animation instead of just continuously spinning like it's supposed to and the piston continually going up and down and then cling just kind of sort of went off the rails completely.

Uh so yeah, that's

10:21

not quite what I'm looking for. So at the moment I'm kind of back to the drawing board on the best way to get the visuals.

We have some visuals we can use from our inv video. So maybe we just start there.

10:37

This time let's use only generated clips. So let's go ahead and generate this.

And now we'll have a couple places to pull from. And then here's the generated video that it made for us.

So I'm going to mute it because we don't really need the audio. Uh but if we skim through it, you can see that it tried to

10:53

generate visuals to go along with it. But these visuals don't really make sense.

They're not accurate to what I'm talking about. So, if I put these in a science explainer video, people will not understand the science from these visuals, unfortunately.

And there's

11:09

really I mean, I've skimmed this whole thing and there's really no good visuals that I can use cuz all of it is just sort of nonsense in relation to what I'm talking about. So, what do I do when I get hung up on a problem and I can't figure out the solution?

Well, I ask AI. And so, I went to Perplexity and I said,

11:25

"What are the best AI tools to find B-roll for a script? I don't need it to be generated videos.

B-roll works just fine. In fact, the videos that we looked at at the very beginning where they were making those music science explainers.

They were using B-roll throughout the whole videos. It was just footage they

11:40

collected from around the internet. Most of it didn't seem generated.

So, B-roll should do the trick. And it suggested Opus Clip, Submagic, Capwing, Hypernatural, V, Big View, Visla.

And I've heard of all of these tools. Most of them I don't think are what I'm

11:56

looking for, but OpusClip might be the solution I'm looking for. I've used Opus Clip a lot and they've been a sponsor on this channel in the past.

Really, really good company. I think it's a great product.

Now, I do see on the Opus Clip web page,

12:12

they do offer AI B-roll. And I can see in my dashboard there is an AI B-roll.

So, it is AI generated B-roll, it looks like, but it wants a video. So, here's what I'm going to do.

I'm going to jump over to Cap Cut and we can see our video has a lot of like black gaps in it where

12:28

there's no video footage. I'm just going to export what we have and there'll be some sections of just blackness.

But at least I've got a video that I can upload and it can try to analyze and find B-roll for. So, let's go ahead and export this.

Click on upload here for AI B-roll.

12:44

Oo, so it gives me an option for AI generated B-roll and stock B-roll. That's perfect.

I mean, maybe we can run through both and see what it gives us. All right.

So, here's our B-roll version, which if I skim through it, there's some usable stuff, maybe, but it's not super relevant either. The

13:01

exhaust valve opens and it's showing what appears to be a lathe. No, that's some part of a machine, but not not what we're talking about in this video.

And then right here where it says the piston rises, it shows somebody swimming, compressing air and fuel tightly. I I

13:18

don't know why. Why?

Why? Why?

Why is that the B-roll it chose? All right, let's look at the generated version.

Fuel is sprayed into the cylinder, but there is some new footage in this

13:33

generated version. Air enters through the intake valve.

Shows a picture of an engine, not super relevant to what's being said. All right, so I think I've come to the conclusion that AI is just not quite there yet to generate all of the shots for a video like this.

We can

13:49

have it generate some of them. We could have Invido and Opus clip and tools like that go help us round up some B-roll.

But at the end of the day, it doesn't seem like AI is really good at creating the imagery for these explainer videos yet. And if we look at the videos we talked about at the very very beginning

14:05

of this video, that was all stock B-roll footage that wasn't generated footage. Hopefully in, you know, 6 months from now, we'll be able to generate the exact footage we're looking for.

It's just not quite there yet. So, to get this first video across the finish line, I'm going to need to source some more stock

14:21

footage. A quick look through my favorite royalty-free sites, and I'm able to pull everything I need to make a compelling explainer video.

From there, I throw everything into Cap Cut, use their auto captioning feature, and hit export on my first viral masterpiece.

14:38

Fuel is sprayed into the cylinder. A piston moves down.

We're drawing the mixture in. The intake valve closes, the piston rises, compress.

14:54

I mean, I think it came out pretty good. You could watch that video and listen to the song and see the visuals and kind of know how a engine works.

But that's not good enough for me. I am dead set on figuring out how to better automate this.

15:13

It may not be as good as what we just saw, but I want to get pretty dang close. So, let's build an automation for this.

I'm going to play with a tool called Glyph. It's a tool I haven't spent a lot of time with, but I flew out to New York and met the Glyph team a couple weeks ago, and they showed off

15:28

their new agent mode while I was visiting with them, and I was pretty impressed. And I think I might be able to get most of what I want done just by building an agent inside of Glyph.

So, let's try it. So once you're inside of Glyph, we can come over to agents here, and we can actually see a ton of

15:45

different agents that have already been pre-built. There might even be one that's already pretty close.

Let's see. Music video creator.

Generate infinitely long one-take videos with Cling 2.5. Selfie music studio.

Make yourself rap, sing, dance. But you know what?

I'm just

16:00

going to build one from scratch since I know exactly what I'm looking for. So let's create an agent.

Once inside Glyph, I choose the chat engine and run through all the instructions it needs to oneshot this task. This includes telling it the order in which to run through the tasks and more importantly which

16:15

platforms and tools to use. And I'm choosing everything.

I'm thinking that if I cast a wide enough net, I'm bound to catch a nice fish. Okay, so we're actually on to something here.

So, it used Clling to generate this video here. It added subtitles and then it overlaid

16:30

the song. All I did was give it this original prompt and it did all of this stuff for us.

Generated the music, generated some images. This is the actual song which I'll play in just a second.

Generated the starting frame, generating an ending frame, animating

16:45

the starting frame to the ending frame, adding the subtitles, and then overlaying the music on top of it. And here's what we got.

Now, it only generated 10 seconds, and the visuals aren't great, but it's strung together a lot of these steps really, really well. That's what I'm impressed by.

17:02

On a circuit board, tiny city of light, copper traces our road, sending signals just right. Resistors slow it down the faster.

Smooth the flow of the chip. Is the brain saying, "Ready, set, go." But it does ask me at the bottom, want me to make a longer 30 to 45se secondond version with a chorus and more scenes?

17:18

Hell yeah, I do. Okay, so we got a 44 second song.

My instructions were not to use V3, and it killed a crap ton of credits, as we can see down here. And then it went to add subtitles to the video and it says it's a 44 second

17:33

video. I'm skeptical.

On a circuit board, tiny city of light. Copper trace is a roads guiding currents in flight.

Power needs a reg.

17:53

All right. I guess I was right to be skeptical on that one.

[Music] Okay, I think I know how I'm going to do this now. I'm going to use Mine Studio.

I think I can build an automation like a

18:09

workflow that with all the steps using Mine Studio. Let's see if we could do it here.

So, build new agent. Now, I've used Mind Studio a lot in the past for setting up these kinds of agentic automations.

It's really clean and pretty easy to use. So, I add each step of the process starting with making the

18:25

song. Take a user inputed topic, research, generate the lyrics, and turn that into a piece of art.

Not slop art. Please let this one be art.

It seemingly worked. Let's listen to a snippet here.

Flat insulating boards with copper veins, fiberglass core for strength and

18:41

contains guide the current flow. Not a huge fan of the style of the song.

So, let's go to our text to speech our 11 labs. If I was to give it more instructions, like create a song, create

18:56

a kids song that's poppy and catchy with a female singer's voice. All right, here's what we got.

Amazing. [Music]

19:14

I kind of got it working by bringing this down to 900 characters. 2000 was still a little bit too long of a song.

And then on the text to speech model, change the length to 30 seconds instead of 45 just for the sake of speed. A

19:29

circuit board is a rigid base. All right, now comes the hardest part, and that's trying to get the visuals to go over the song itself.

So, I'm starting with V3 fast in order to save some time and money on the generations.

19:44

But I'm feeling like my frugality may come back to haunt me. I'm also thinking that if I split up the lyrics into chunks, I should be able to get separate clips for each part of the song.

I think then I'm going to use Mine Studio to automate the assembly of each

20:01

component of the video. Couple of final touches.

And here's what it generated for us. A 30 second video with a song, not as good as Sudo's song, and it did put visuals throughout the whole thing.

[Music]

20:18

A circuit board BCP holds parts and links them links them. Fiberglass carries thin copper traces in

20:35

layers and holes where it did all the steps. I just can't get the visuals to look the way I want them.

The choice is going to be you you get all the way up to where you have the

20:50

song created and then you go in, pull the visuals and sort of splice them all in where you want them or you can get a fully automated music video that's generated with AI. The visuals just aren't going to be amazing if you're trying to get it to be like that

21:06

explaining how something works. There are ways to do it with AI.

Whether those ways are faster than just sourcing B-roll right now, I I don't think so. But yeah, for this style of music video, you could do a big bulk of it with AI,

21:23

but at some point you're going to have to edit the video together yourself or hire somebody to edit the video because I just I don't see a way to just generate that whole process yet. Hopefully, you did learn something.

I mean, we did figure out how to make that style of video. It just wasn't all AI.

21:40

had to source some of the videos. And then we also figured out how to use some pretty cool automation tools.

Although we didn't get our end result, I have a much better idea of how to use glyph now and some cool workflows in Mind Studio that I didn't know I can do before. So,

21:55

still learned a lot along the way. I guess at the end of the day, if this got your wheels turning about some things that you can create, you know, maybe it's not trying to make educational music videos with uh perfect B-roll to go along with the music.

Maybe it's, you know, Facebook video ads for your

22:12

company or a silly meme generator that generates brain rot slop. It'll probably be pretty good at that if that's what you're going for.

But to get that educational style video, still a little bit of work involved, meaning we're probably not going to see the internet flooded with them quite yet. But that's

22:28

what I got for you. Thank you so much for tuning in.

If you like videos like this and you want to see me dig deeper and explore a lot of other AI tools, make sure you like this video, subscribe to this channel. I'll make sure more stuff shows up like this in your feed.

Really appreciate you hanging out, nerding out, and hopefully I'll see you

22:44

in the next one. Bye-bye.

A circuit board BCP holds parts and links them links there. Fiberglass carries thin copper traces in

23:02

layers, pads and holes, marks

Summary

Transcript