WEBVTT

00:00.000 --> 00:03.250
Welcome to AI for visual effects
I'm Doug Hogan.

00:03.916 --> 00:06.333
In this video,
I want to show you how AI filmmaking

00:06.333 --> 00:09.791
actually works in production,
because the reality is

00:09.791 --> 00:13.375
you almost never use
just one model from start to finish.

00:13.416 --> 00:16.666
Instead, it's a mix and match process.

00:16.708 --> 00:19.666
Different models are better
at different kinds of shots,

00:19.666 --> 00:22.625
and this setup is designed
to take advantage of that.

00:22.625 --> 00:25.166
We're going to start
by defining our shot.

00:25.166 --> 00:30.666
Up top. you'll enter two simple things
what the image is and what the action is.

00:30.750 --> 00:33.041
Think of this
like you're briefing your DP.

00:33.041 --> 00:37.666
This info gets passed along into an LLM
with special system instructions

00:37.666 --> 00:41.125
to build a clean cinematic prompt
for us automatically.

00:41.125 --> 00:45.875
So instead of handwriting prompts,
we're using AI to direct AI.

00:46.125 --> 00:49.416
From there,
we generate our initial steel frame.

00:49.416 --> 00:51.375
This is our hero frame.

00:51.375 --> 00:55.291
It defines the subject, the lighting,
and the overall look of the shot.

00:55.291 --> 00:59.291
And this step is important
because everything downstream is going to

00:59.333 --> 01:00.875
build off of this image.

01:00.875 --> 01:03.291
Next we move into camera control.

01:03.291 --> 01:08.875
Here we use a second model to reinterpret
the image and generate camera aware variations.

01:08.875 --> 01:11.875
This is where you can subtly adjust
framing, angle

01:11.875 --> 01:15.208
and composition
without breaking the identity of the shot.

01:15.208 --> 01:16.583
Think of this like doing layout

01:16.583 --> 01:19.833
or previs adjustments
before you commit to final rendering.

01:19.833 --> 01:22.708
Small changes
here can make a big difference later.

01:22.708 --> 01:25.666
Once you're happy with your frame,
we lock it in.

01:25.666 --> 01:29.000
That image becomes the source
for everything that follows.

01:29.041 --> 01:32.041
Now this is where the workflow
really opens up.

01:32.083 --> 01:33.250
We take that single

01:33.250 --> 01:37.333
image and shot concept
and push it through multiple video models.

01:37.541 --> 01:49.333
VEO 3.1, Grok Imagine, Kling 3.0,
Seedance 2.0, LTX 2.3 and Wan 2.2.

01:49.333 --> 01:53.250
Each one gets the same shot,
but each interprets motion, timing,

01:53.250 --> 01:57.708
and realism differently with the help
of another LLM with special system

01:57.708 --> 02:01.958
instructions to tailor our prompt
for each model specific preferences.

02:01.958 --> 02:03.916
And this is the key idea.

02:03.916 --> 02:06.083
You're not trying to find
the perfect model.

02:06.083 --> 02:09.833
You're generating multiple
takes and choosing which one works best.

02:09.875 --> 02:14.000
Just like filming a scene on set,
one model might give you better motion

02:14.000 --> 02:18.000
for this shot, while another might give
you better lighting for that shot.

02:18.166 --> 02:22.208
You generate a range of options,
review them, and pick the best one.

02:22.291 --> 02:24.041
A couple quick tips here:

02:24.041 --> 02:27.375
First, don't hit play
until you're happy with your frame.

02:27.375 --> 02:31.250
Some of these video nodes can use credits,
so you want to lock your composition

02:31.250 --> 02:32.958
first before wasting them.

02:32.958 --> 02:37.625
Second, if something feels off in the
final result, go back to the image stage.

02:37.875 --> 02:42.750
Nine times out of ten, fixing the initial
start frame fixes the entire shot.

02:42.833 --> 02:44.500
So have some fun and see that

02:44.500 --> 02:48.250
AI filmmaking isn't
about using one model to rule them all.

02:48.291 --> 02:51.958
It's about using the right model
for the right kind of the shot.

02:52.083 --> 02:56.458
And this workflow gives you a fast,
flexible way to do exactly that.