With today's video models, you can do

whatever you want. Like this. Want to

learn how? Let me show you. So, this

video pipeline is nuts, and I'm really

surprised more people aren't using

something like this. Um this is

extremely powerful and hooks, like you

just saw. It's also really good in ad

creative, as well as organic stuff. Uh I

have a friend who is worth many tens of

millions of dollars that's currently

generating over 2,000 AI videos a week

with something like this. So, it's not

just changing an outfit. I have kind of

this like little USB thing. Um well, I

can do it just like this. Alternatively,

let's say I wanted to, I don't know,

change the time of day. Um re-lighting

and stuff like that is pretty expensive

if you think about typical productions.

Now, I have the ability to change

lighting in a flash. So, I'm going to

show you all that in this video. You're

going to learn everything you need in

order to be able to develop really cool

visuals like this. So, what is the

actual workflow or like pipeline for

this? Well, it's pretty straightforward.

To start, you need some form of source

video. And the source video in our case

is going to be what you guys just saw me

show you earlier. So, if I just head

over to finder here, you can see that I

basically have this folder called AI

video set up. And inside of that folder,

if I just double click, we have that

same video. So, if I just go through

with no audio, you could see that I lean

forward and do the exact same thing that

I did in that intro, and then I even

snap my finger. It's just notice that

when I do the actual snap, nothing

happens. Well, the reason why is because

we're going to use this as the trigger

to change my outfit. Now, that source

footage can be around 10 seconds or so.

I don't think most video models right

now allow you to go over this. You can

also creatively chain together a bunch

of these source videos in a way I'm

going to show you guys with no downside

in a sec. After you have that source

video, you need some sort of hyper

specific prompt. Now, the way that I

define a hyper specific prompt, and I've

just seen a lot of people screw the

pooch on this one, is you just need some

form of trigger, which is typically

time-gated. It's like a thing that that

occurs during the video that results in

a change. And so, in the demo that you

guys saw earlier, for instance, it was

the snap, right? The snap did the

change. You can tie it to specific words

and so on and so forth, too. So, it's

really not that complicated, right? You

have a trigger over here, which is the

moment that the model watches for. And

then you have a change, which is exactly

what is meant to happen next. So, in

this hypothetical example where I'm like

wearing a cloak or something, when the

man reaches towards his shoulders, maybe

like this, have them pull a shimmering

cloak over himself and fade to

invisible, maybe Harry Potter style. So,

if you want to have a good prompt that

works, if you want this to work, you'll

notice that the specificity is very,

very important. And I'll make sure to

show you guys a bunch of examples of

that in a sec. Okay, finally, you

obviously need the AI intelligence. Now,

there are a variety of different models

currently available that you could use.

I'm going to be using one called Omni,

but Kling is also pretty dope. And as

mentioned, people are developing these

things all the time. The key with the

sort of model to use for this is it

needs to be video to video, which is

different from like your typical

run-of-the-mill video generators. Uh,

and let me cover what that means. Right

now, a lot of people here that have

experimented with like AI video

typically get pretty disappointing

results. And the reason why is because

most of the time, you're trying to

generate something from scratch. AKA,

you are just feeding in text, and then

you're expecting a bunch of pixels in

the time dimension to magically appear

and to be totally coherent.

As I'm sure you can imagine, this is

pretty like intellectually complicated.

I mean, you can't I can't do that. Takes

me a hell of a lot longer than like 30

seconds or however long these video

models are. So, rather than generate

from total scratch, okay, and then be

disappointed with the results, we're

working in a medium that we've already

recorded. And so, we we provide some

sort of real footage here. And then all

we do is we actually just modify that

footage ever so slightly with a prompt.

Um, and so that's sort of what we're

going to do over here with Omni. We're

actually going to like feed in that the

pixels, the lighting, everything's

already going to be there. All we need

to do is just basically insert something

into this world that we've already

created. Obviously, inserting something

into a world that we've already created

is a lot easier than creating an

entirely new world. Makes sense. Now,

finally, when all is said and done, um,

we're going to output some sort of 720p

render. For those of you guys that are

unfamiliar, most video nowadays,

especially video on YouTube and other

social platforms, is higher than 720p.

And the reason why I bring this up is

because the people that build AI video

into their workflows that don't account

for the fact that the quality is

significantly degraded, at least as of

the time of this recording, I'm sure

it'll get better eventually. The people

that don't learn how to work around this

fact are going to get significantly

crappier results and their pipeline is

just going to be terrible. What you need

to do in order to make this work for any

ad, any organic video, whatever, is you

need to hide the 720p seam with an

intelligent cut to a different scene.

Now, what do I mean? I mean, if you guys

have a crappy AI shot at 720p or, you

know, like a more compressed version and

then you immediately cut back to a real

shot, one that has not been AI edited,

it'll be naturally disjointed. Um, a

really good example of that is like

this. I mean, this just looks really

weird all of the sudden because

obviously, you know, it gotten a lot

more compressed. So, what that means is

instead, you need to change the actual

nature of the shot. Rather than going

from like full screen talking face to

full screen talking face, you need to go

from like full screen talking face to

just another scene entirely. Um, the way

that I do it and the way that I've done

it many times in this video, if you

haven't seen, is I go from the 720p AI

shot to a screen share. Can have some

super crazy happen in the

background, but then cut back to the

video, uh, you know, you have no idea

that that occurred. You can't really

tell the pixels are different. Nobody

can really say, "Hey, you know, what

happened with that weird scene?" And

that's it. Now that you understand

everything that goes into this, let me

show you how to actually do it

yourselves. Now, you need that video

model like I talked about. The first one

I'm going to be using is Gemini Omni.

And if you head over to

deepmind.google/models/gemini-omni,

um, you can go to this page right here.

You can try it directly in Gemini, also

Google Flow, and then they have a little

build with Gemini Omni feature down

below. And you know, you guys can see

the scope of changes you're capable of

making. You could turn your whole room

into bubbles if you wanted. Really, it's

more about the ideas than it is anything

else. Um, you just need to like know

what to tell the model in order to make

it work the way that you want. This guy

turned himself into like Neo from The

Matrix, basically.

Um, now, I'll be honest, it's not free,

obviously. We're going to be doing

something that's pretty computationally

intense. Because we're going to be doing

something that's pretty computationally

intense, you guys can expect it to cost

a little bit of money. Um we're working

with pixels here and not just static

pixels like with images, we're working

with a time dimension. Keep in mind that

on average a video is 24 to 30 frames a

second. So you're technically generating

24 to 30 images a second. We had 8 to 10

second video clips which I'm going to

show you guys in a second. You know, one

of these things might realistically cost

you 50 cents in order to generate. And

what you'll see is you can't just

generate one, you typically have to

generate multiple simultaneously to take

advantage of parallelism and then

exploring a vast solution space. More on

that in a sec. Now because of this,

because of the fact that it's expensive

and then I got to come up with all these

prompts and stuff like that on my own

and to be honest, I'm not a very

creative person. I don't actually use um

Gemini Omni directly in their builder. I

use it through like a third-party

platform, in my case Higgsfield. There

are a couple of these but essentially

what they are are model aggregators that

just slam together all available current

video models, state-of-the-art systems

and stuff like that, just into one place

so that you can build a pipeline where

you pass it through, let's say model A

first and then go to like model B after

and then do model C after that. And you

know, with so many different models out

there, this is really their value prop.

It's like, oh you know, just come to us

and we'll deal with all of it. So I like

Higgsfield. I'm going to use it just to

show you guys more or less how this

stuff works. But I want you to know you

don't have to be tied to this. You can

also just use it in the Google Omni sort

of dashboard. When you sign up to this

and then you click on video in the top

left-hand corner and that's what I would

do. Don't click on a specific one, just

go video. You'll be taken to a page that

looks something like this. If you head

to the top left-hand corner, click the

button, you'll get something that looks

like this which is just like a big list

of prompts that you can use to one-shot.

I'm not going to do any of that right

now but it's pretty neat. You can just

like select one, um have a really

engaging intro hook for your ad or

whatever and then immediately arbitrage

tokens and then like the increased CPCs.

It's actually kind of nuts how you can

do that nowadays.

In terms of the actual workflow, it's

pretty straightforward. So I'm just

going to exit out of this, pretend I

haven't actually uploaded my media yet.

Um I'll head over to upload media and

you'll see here that I have like a video

combined 2026 one that I was talking

about um, that I'm going to have

uploaded. After it's done this little

uploading dialogue, all you need to do

is just click on it, and then that'll uh

put it right over here in the left-hand

side, added to prompt box, and then just

enter whatever it is that you want. And

in my case, what I'm going to do with my

hyper-specific prompt is I'll say,

"Right after the man says this at

exactly 2.9 seconds, change his outfit

to a cool-looking hoodie with a chain."

Then I'm going to click generate.

Now, one thing I want to stop and

disvalue of right now if I'm thinking is

that this is going to work 100% of the

time. This is going to work like 20% of

the time, which means on average you're

probably going to need to rerun this

thing five times or so if you really

want to get the smoothest effect

possible. Nobody's telling you this

right now because they want you to think

that this is way cheaper than it

actually is, um, but you're going to

have to run this multiple times. So,

just knowing right off the bat that I'm

probably going to have to run it

multiple times, rather than just like

make a prompt, click generate, wait,

wait for the results, see the results,

be dissatisfied with the results, and do

it all over again, rather than spend

like 30 seconds on that, I'm actually

just going to generate a bunch of these

simultaneously.

And I'm going to do five or so, and I'm

just going to see which one is the best.

Um, this also takes a fair amount of

time cuz as mentioned, we're not just

doing, you know, like one image, even

images take a lot of time. We're doing

like 24 to 30 images a second for

however long the length of the video is.

And you'll [snorts] also notice that

this isn't free. I mean, this costs 15

tokens. I don't know exactly what the

token to dollar conversion is here, but

you know, if you do this four or five

times, that's like 60, 70 tokens. That

can actually add up a fair amount. You

should expect to spend maybe 50 cents to

a dollar for this. Obviously, as time

goes on, this is going to go way

cheaper, but, um, some ways that you can

significantly reduce the cost if you

guys are cost-bottlenecked, is you can

upload a shorter video. Uh, if your

video is like 10 seconds, you're going

to consume more than 15. If it's like 5

seconds, you're going to consume less

than 15. Um, I'm doing 16 by 9 here

because that's wide screen, uh, but

obviously, you know, you guys can select

the smallest aspect ratio for whatever

the specific model is that you want. And

then, yeah, I mean, like the the shorter

the video, uh experiment with different

types of models if, you know, cost the

major bottleneck. And then you can

eventually get to the point where I'm

going to show you here. After all said

and done, you get something like this.

This just generated. I'm just going to

turn on my audio. Hopefully, you guys

can hear this.

With today's video models, you can do

whatever you want. Like this. Want to

learn how? Let me show you. And you'll

see that, you know, it hallucinate from

time to time. Like what happened there?

I didn't even snap my fingers and then

it changed my outfit. And then, not only

did it change my outfit once, it changed

my outfit again. And then it changed my

outfit again to like a half blue suit,

right? So, that's not actually what I

want. I just want like the the super

finesse chain switch. So, I'm going to

do the same thing here.

With today's video models, you can do

whatever you want. Like this. Want to

learn how? Let me show you.

With today's video models, you can do

This did two things. I mean, like that I

didn't really like. The first thing is

it sort of like slowly

like spanned the hoodie over me, which

is actually kind of dope if you go frame

by frame. But, um then it like put the

mic in my hands, which I don't really

like. I just wanted it to look exactly

like you guys expected.

With today's video models, you can do

whatever you want. Like this. Want to

learn how? Let me show you. This one was

today's video

This one was pretty good. I think this

is probably like 80 90% of the way

there. I could probably select this and

move on. I'm just going to poke around

and see if the other couple of gens that

I made are okay and then I'll loop back.

Okay, and then the end output, the thing

that I liked actually ended up looking

like this. With today's video models,

you can do whatever you want. Like this.

Want to learn how? Let me show you. So,

as you guys have already seen this cuz I

put this in the intro. Um that snap, it

happened immediately, that millisecond

that I did it. We have to do is we have

to weave this into whatever other

footage that we're doing with the scene

change. So, you know, in this case, I'm

going to use Premiere Pro. Uh you guys

don't have to be video editor pros or

anything like that. Uh Premiere Pro also

costs money like, you know, Higgs Field

like Omni and stuff like that. Um so, to

be clear, if you want to work with AI

video, you know, it's not going to be

cheap. It's going to cost you quite a

pretty penny.

Uh I just like contrast that to like

token-based pricing, right? But anyway,

I'm just going to open up Premiere Pro

and then I'm going to feed in my

content. So, I'm going to go shorter

over here. It's just like the name of my

project template for whatever reason.

I'm going to open that puppy up.

And then what I'll do, and keep in mind,

you can do this again in whatever thing

you want. Okay, I'm going to feed in the

original footage.

And we're just going to keep existing

settings. Then I'm also going to feed in

this Higgs Field edited footage. So,

just so you guys could see the actual

difference.

Okay, so this is the high-quality intro.

And let me just make sure that it's on

full. Do you guys see how like the

pixels are really defined and stuff like

that?

This is the Higgs Field 720p.

And so, I mean like if we kind of, I

don't know, hypothetically were to

overlay this one on top of each other,

like this. And then if I were to go back

here, okay, and then play, you'll also

notice that the length of the videos are

just a little bit different. Do you see

how this one here goes to here, whereas

this one here is sort of different?

That's because the AI is actually

pretty, like it's kind of changed the

very nature of the video itself,

including like the audio waveform and

when things are in the in the audio

file. So, you know, what I'm going to do

is I'm just going to very try very

carefully try and drag this over so it's

like one-to-one. And then at any point

in the video,

I'm just going to go back and forth so

you guys could see. So, this is the Omni

one. This is the real one. Omni, real.

Omni, real. You can see there are very

slight differences, but we're nowhere

near like uncanny valley territory like

we are when, you know, we realistically

try and AI generate a totally new video

clip. Um, so now that you have that,

what you need to do is you need to weave

that into like a piece of footage where,

uh, you know, we're not actually

transitioning right back to another

talking head screen cuz if we do that,

it'll kind of be broken. Okay, now I'm

just going to find a really simple video

I could use as an example. Why don't we

use this one here?

This one looks like it's kind of cutting

into a bunch of stuff, so that looks

nice.

And I'll go back here. All right. Okay,

so now what we have, I just need this

track cuz we have this sort of magical,

you know, snap my finger thing. And then

notice how I'm transitioning to a new

scene.

And so, you can't actually tell that we

just went from like AI to something

else. And obviously ideally whatever the

footage is that you have if you're doing

like a screen share talking head thing,

that would be pretty similar to.

Okay, so that's more or less it in a

nutshell. The cool thing is you can

actually proceduralize this with Claude

if you guys are using Claude code or a

codex if you guys are using codex. Um

the way that you do that is at least

Text Field anyway has like an MCP which

is just a series of eight API connectors

essentially that you can call that can

do all of this stuff for you. So you

could actually say like yo, I got this

thing on my computer, you know, I want

to edit it 10 times and then I'm just

going to select the best one. And that's

pretty much it. If you guys like this

sort of video, let me know down below.

As mentioned, the key here is you need

to be pretty specific with your prompt.

So ideally you'd either denote the exact

moment that you want something to happen

or the exact trigger condition with the

change. You also need to retry a bunch

of times. Like this isn't going to

happen immediately for you. That's just

the state of AI video.

That said, if you guys do get to a point

where the pipeline works really well for

you, whatever the effect that you're

adding and and so on and so forth is,

this is extraordinarily effective right

now. Like you can arbitrage the credits

that you spent on this to like massively

improve, you know, your ad CPCs, you

know, the the watch time and retention

of your content in specific places and

so on and so forth. And basically

nobody's doing it right now which is

wild. So yeah, let me know down below if

you guys like that sort of stuff. I'll

make tons more AI video content for you

if so. Also if you guys want all the

resources from this video, I've actually

uploaded them in the classroom of Maker

Zero. It's my free school community

where you essentially just upload

everything to make it really easy and

centralized. At the same time there are

a bunch of additional benefits including

comprehensive courses, good prompts and

and workflows and loops and stuff like

that. So if you want that, just head

over to Maker Zero, click classroom, and

then head over to recently uploaded for

everything you need from today's

session. Thank you very much for

watching and have a lovely rest of the

day.