I said I'm getting nerdier just watching
you.

Look at all these terms. Distilled.

Ooh.

Ha

Alright, David. In the last episode we had
a pretty funny intro.

Do you want to show us how you did that?

Yeah, it was really straightforward.

Alright, so the way that I did it was
using LTX 2.3,

~ which is made by Lightricks out of
Jerusalem.

And so thing is with the system
requirements of the base model,

as you'll notice here, is that it requires
a lot of VRAM.

And so how did I get that to work?

I don't have ~ this enormous ~ GPU.

So the way that I did it was ~ using a
quantized version.

So ~ FP8 means eight bit floating point,

doesn't matter, but that is the version of
the model that I used.

And let me show you the workflow.

Alright, so in Comfy UI, come over here to
templates and then search for LTX.

And then you're gonna see this workflow
over here for first last frame to video.

That's what FLF2V means.

So when I took me a little bit.

How long did it take you to figure that
out?

It's all technical terms. So ~ when you
first open this workflow,

you're gonna see some errors maybe.

and that's because you might not have the
~ files specified ~

and the models downloaded. So come over
here to the right hand side where

it says see errors. And there is a
download all button that you can

use to just download all of the models
that you need.

You can see here that this is using the
8-bit floating point version of

LTX 2.3 and it's using another quantized
version of ~ Gemma 3

Alright, so once you have that downloaded,

you you select it over here, click on the
check mark,

and you're good to go.

And now all that you need to do is choose
your last and first frame.

Alright, so I've selected my two images,

click on the check marks, and look,

it's it's me. It's me in my and my chair.

~ so here's an image of the chair without
me and one with me.

So over here, as per normal, there are
these nodes for us to work with

and just a few parameters for us to tinker
with.

So first what I need to do is make sure
that the height and width match my image,

or at least it's the dimensions that I
want for my output,

so that's fine. 720 x 1280 the frames per
second we don't need to touch,

same thing with everything else.

now we need to choose how long we want
this video to be.

Five seconds is fine.

How long can you make it? Is this ~ one of
those infinite talk models where

you can make like a two minute video?

I don't know. Why don't we go try it out
after we make sure that it succeeds?

All right, so ~ now all that's left is
giving it a prompt.

So let's say in a flash of light,

the man appears.

All right, on it goes.

I gotta say, this is a pretty simple
workflow compared to some of

the ones that you've shown, David.

It is pretty straightforward, isn't it?

First image, last image, prompt.

That's it.

Now you can bring this up at ~ all of the
parties that you're at to talk about

~ FLF2V. ⁓

That's right.

Yeah, I always make sure to bring up ~
distillation and quantizing when

I'm at cocktail parties.

Alright, so this five second video took
five minutes to generate locally

on my machine with ~ you know,

just retail hardware. Let's see how it
did.

~ well, the I don't see a flash of light.

But hm had to think about that.

Mm.

Well that was ~ interesting. ~ I mean,

it made the transition.

Mm-hmm. Is this a prompting issue?

Like is this something where if you used a
notebook LM with good prompting

instructions, then it might have done a
better job.

Or is this a just increase the number of
turns?

I wonder whether it's ~ just the seed.

and if you notice here there's a noise
seed and anybody who's familiar with

how computers randomize, you start with a
seed.

basically I think it might be a random
chance that it didn't do

any flashing of light. So let's give it
another roll.

C'mon seven.

Ha ha ha.

that was quick.

So ~ I guess the the system's warmed up
now because that five seconds video took

a little bit over a minute to create.

So let's see what it did.

Okay. I think it it just it really
overindexes on on the the man appears.

~ so why don't we try a different prompt?

Alright, so I went into Claude ⁓ and
told it to make this a better LTX 2.3
prompt.

here's what I came up with. there's a lot
that it kind of inferred here.

So I'm gonna tweak this a little bit

I think that should be fine. All right,

let's take claude prompt and go.

Alright, so it's done making the video.

Let's see what we've got.

There we go.

There we go. Something like that.

~ I don't know about the whole blurring
effect,

but hey, we got something. so I think the
lesson here is just keep re-rolling

and also get a better prompt, maybe
generated by your LLM of choice.

That's awesome. Well well done,

LTX,

and thank you for showing us that,

David. This was super cool.

I guess we know what the intro will be for
our next video.

Yes, exactly. All flashes of light.

Alright, with that you got to see how to
create a video with LTX 2.3,

just with the first frame and last frame.

And if you want to get more great content
on how to use AI

and agents for your work, follow us at
Prompt and Circumstance.

All right. We'll see you next time.

See you then.