Hi guys. Hi everyone. Thank you to be

here again with me with Edo in this

channel. I want to thank you for your

support, for your feedback, for

everything you ask me in the comments.

I'm It's a pleasure to me. Um today I'm

I want to make a part two video, okay?

um of the video with LTX 2.3 where we

saw how to install how to surfing around

this you know workflow simple workflow

for one image for one image

uh to have something uh more specific uh

with you know um just prompt and first

frame. Today I want to show you

something different.

This workflow here uh useful to create

something with first and last frame

video. Uh it's powerful and I think it's

very cool. Um

and thanks to what dream cost confi

nodes. Um I tried many many many many of

these first last frame workflow and no

one worked. No one no single one worked

well.

This is the only one worked for me.

Worked well. Okay. So I just create I

can create a real first last frame video

in local. Remember in local with my 5060

Ti 16 gig of RAM and 74 gig of RAM in my

spec. I'm working with the Ryzen 9. So

this worked for me.

I hope will work for you as well. Okay.

Uh the workflow it's quite simple. It's

quite clever. Again compliment again was

the dream cost. Very very very cool. uh

workflow. Um, and I want to show you how

to work with this workflow.

Uh, he made um he made also uh you know

um

um a a

video a YouTube video you can find in

your uh in this uh tutorial if you want

to follow him the the the creator. I

want to show you my my my my opinion and

my tests. Okay. Um as you can see we

have three parts of the the the you know

the model. The first part is the loader

with everything in here. Okay. The

second part is the multi- image loader

created to uh to have the control of the

images. And then we have the upscale

part.

Okay.

With two two stage of us upscaling if

you want. I'm using just one. Okay. So

the the idea it's to insert two you know

two images the first and the last

uh create um a sequencer.

Okay. If I remove everything, it will

remove also the images. But I want to

insert again. Okay. And it will create

this. We work with frames. Uh and we

saying to the model, insert the first

images here. Okay. Uh and the last frame

is the second one. If I'm if I die the

deal minus one as here. Okay. I tell to

the model this is the last frame so work

as you want. By the way, we can work

with the seconds

and we want to insert the second frame

when we want in terms of seconds and

work with frames. So uh you know I'm

just leave this and we

have to deal with something different

from the previous one model. As you can

remember this model here we have just

one powerful model or the checkpoint on

one Lora. Everything else is in this

model here. For this workflow,

we need a different

loader.

And for this workflow, you have to

update your KJ confi nodes. How to do

this? Right here, right here, and

update.

If you don't update these notes here,

this workflow will not work. Okay. So

update everything everything and then

use the the last checkpoint we used for

LTX. Okay, I'm using the dev FB8

less powerful but very fast. And we have

to download this

um v

here. This they here and this also here

you can find in your uh right there uh

you can find here here and here there

are different um

different this is the the the workflow

for for LTX. Okay. This is the uh

upscaler. Okay. Okay. And this is the V

LTX.

Uh,

and we are here. I'm using the Laura I

used previously in the LTX single frame

and the two text encoder Gemma and LTX

checkpoint. And this is the spe special

upscaler special upscaler. Okay. And

here we have the value of seconds. We

can say seven whatsoever. And it's very

clever because he create this node with

a mathemat math um uh you know

operation. Uh it's very very simple but

clever cool because we don't have to

make everything every time the

conversion

uh from uh the number of frame and the

value in seconds.

So it's it's simple but it's cool.

Great, great job.

And and this is everything you need to

work with this uh workflow. Okay. Um

this is everything you need. Yes. I want

to show you something else. I don't

think so. No, it's very simple. Here we

have the process latent right there with

the

you know the upscale that reduce the the

the the

scale of your image the the the rate of

your image by a half and then we'll

upscale again by two. Okay, this is the

stage one

and this is then the stage two here.

Here, right there. I just uh change this

um cfg to ancestral

uh C cfgpp or uler cfgppp.

I read somewhere that it's the best for

LTX, so I work with this. Okay. Uh or

you can just leave as you want. And this

is the the second one uler as CFG. As

you can see that the noise for the

second part.

And leave everything as you see. It's

very very precise the the the flux. So

just leave it. Okay. If you want you can

create another stage of refiner with a

third part but you know for me it's fine

just this okay as you can see another

another another stage of upscale

and that's it.

Okay, this is the the decoder of your of

course and we can see everything right

there with the tile the temporal size

and the overlap.

This is the default value for this

model. You can work with this. I work

with this fine. So why not use this

default values and we can

see the last part of the video uh the

prompt. Remember LTX it's very sensitive

to prompt injection. So try to be very

precise in what you see in the scene and

what you want to see in the movement.

Okay.

In this particular scene, a

photorealistic high quality cinematic

video transition. I want to see the

transition. The woman begins to stand up

gracefully from the stool. So I just

create the perfect you know match with

the the the image as she rise the camera

performance control fluid crane up and

slide up pan in movement to maintain

high levels focus and adjust blah blah

blah simultaneously she turns his body

and face the camera blah blah blah blah

blah blah and she says cheers have a

nice party the background of glowing

leor bottles and warmah

B. So be precise.

Be precise if you want a precise uh

movement and a precise video with LTX.

It's demanding. So uh if you create a

poor um poor

um prompt, the video will be poor. Okay.

Just another thing if you want these two

these two uh you know um these two

values is very important the pro the

compression the lowest value the the the

the detail will be uh more precise

crispies sometimes so balance in your

with your um you know pleasure I'm very

good with 15 20 values

Okay. And then remember the the multiply

this loader will in insert the the

images and crop it in a right way in

order to then upscale again. So in my

opinion it's very important to have a

first frame with the right ratio. Okay.

divide uh that you can divide by 30 32

or 16. Okay, in my case work well this

value that is the standard for LTX

uh se 736

and 1,280

pixels. Okay, this is divided you can

divide it by 32. It's important just to

have the right ratio and the right

dropping ratio with LTX and not to have

something strange or you know messy. So

this is all you need to know to perform

a run with this model and I want to show

you the the the result. Okay,

this is result.

Okay, again

as you can see he moved she moved the

the chair.

Cheers party. It's very good. It's very

very smooth. Okay, I g I made another

one,

but the the the torsion of the head it's

quite quite

uh

it's good but not so good. Okay, maybe

there is another one. I think it's this.

Okay, sorry

I made a mess

from the last video I made. uh you know

I perform a a zoom in from uh this start

video

start frame to this last frame.

This is in local. This is open source

and I take 300 seconds. 300 seconds.

Okay. Five minutes. six minutes with my

6050 Ti. You know this is the value

500 seconds

for you know for six uh length video six

seconds length video or 10 second length

video I think it's impressive very very

impressive I think it's quite good uh

you can find I want to show you

something very strange change happened.

This is one of uh with a a wrong prompt.

I made a mess with the prompt and I want

and I wanted to create something very

long. 10 seconds.

Oh, whoa. What's going on? Come on.

Sorry. Again.

Okay. It crashed everything. I don't

know why.

Okay.

Strange.

Maybe I asked

something wrong to my GPU. I don't know

why. Uh, okay. And another one. Okay.

This is good. This is okay. I don't know

why. This is a 10 seconds. So, she

stands up. She go back and she moved the

chair. Blah blah blah.

Okay.

is not good but it's not so bad. Okay.

In terms of what are we doing with open

source

open source guys.

So I think it's an incredible way to

work with AI in a open source uh uh you

know perspective.

This is the top the top the aex of the

the video generation model

and I hope this

workflow could help you. Thanks again

Wild Minder.

No sorry not wild mind. Thank you also

one minder. Okay. Wet dream cost. What

dream cost? Thank you guy. Thank you

very much. It's Jonathan Jonathan

Watkins. Jonathan, thank you. Thank you

for your uh you know um your your

effort. It's powerful. It's it's very uh

you know clean, beautiful to to see and

to uh to work with. So thanks. Uh it

worked for me. I hope will work for you.

And again, thank you to to be here with

me with Edo and I hope this kind of

video could help you in your AI journey.

I work with this um advanced generation

um modus all days and I hope my effort

could help you in your journey and I

hope to see you again next video.