Hey David, is the quality of your

audio generation going down?

In Suno, people are

certainly feeling that way.

They are leaving Suno because they feel

that version 6 is worse than version 5.

Yeah, it looks like they're saying that

the songs they're getting are

soulless garbage.

Yep, and so that on top

of the fact that, of course,

Suno's involved in some

significant lawsuits,

makes you wonder what

alternatives are there for me.

So, people are upset at Suno, they're

leaving, what can they do about it?

Well, here is a great alternative.

It's open-weights, you can

use it right on your machine.

It's YuE2.

And so this is a collaboration across a

variety of different organizations,

including Stanford, NYU, and the Hong

Kong University of Science and Tech.

The performance of this is

competitive with Suno V5 and V6.

So you can also get soulless

garbage out of this?

You can get soulless

garbage really quickly.

Yes, that's right.

So it's a new model.

I see some very high-profile names

there who have collaborated on this.

What have they figured

out that's different?

So, what's different is that it generates

basically the sheet music first,

and then it uses that as scaffolding

to generate the finished song.

Okay, that's interesting.

So, Suno isn't doing that.

It's just generating the

song directly, I guess.

Well, we don't know that.

It's a closed model, but presumably

it's going straight to the end.

So here's how YuE2

performs versus Suno.

So you can see here, here's YuE, right?

And the further to the right,

the better it is in terms of song quality.

And you can see it outperforming

Suno V5, and then of course,

you know, V6 being worse is

a little bit further to the left.

And in both cases, you have

YuE outperforming this.

So, to use YuE2,

you can come into ComfyUI,

where we do image generating and

video generating, open up the templates

menu, and then come here to audio,

or you can just simply search for YuE2.

You'll see that there are two

workflows you can choose from.

Today, we're just going

to cover text to music.

Music cover is really cool,

but that'll be for another time.

David, I got to say,

this looks shockingly simple

compared to some other

ComfyUI workflows.

It is, right?

There's only two main nodes for you to

work with, and it's really straightforward.

In fact, there's only just one

file for you to download.

So, by default, the workflow wants you to

download this 8-bit quantized version of it,

so that quantized

version is great for those

who have very little VRAM,

like 4 gigs maybe.

But if you have something

like 8 gigs of VRAM

or more, just use this one,

it's better quality.

So, once you have that downloaded,

simply refresh with the R key, right?

So you press the R key and that'll

refresh everything here so that it'll pick

up on the files, and then come here

to select what you've downloaded.

Make sure, of course, that you

downloaded it into the checkpoints folder.

So next, you need to

provide it with a prompt

on what kind of song

you want to generate.

And then also provide the lyrics.

And in the lyrics, you can say,

all right, this part

is the verse, this part is

the chorus, and so on.

Won't you get somebody singing chorus?

No, you won't.

In fact, it's pretty good at recognizing that

things in square brackets

don't need to be said.

And where are the sliders for how

soulless or generic you want it to be?

There aren't any, thankfully.

The only other thing that you need

to tinker with is this max duration.

So by default, it comes with

just two minutes max duration.

I give it a bigger number, right?

So 400 seconds is like six

minutes and change, which is

more than what you would

probably want for your music.

So if you want to regenerate this,

you can certainly do that.

However, you got to notice that the seed

here is what determines what gets made.

And here, you got to manually do that.

And I don't like to manually

do stuff with the ComfyUI.

So come into here,

into the text to music node,

and then scroll down to the

seed node and just unplug it.

Just unplug that, and then come

here and choose randomize.

What'll happen now is

that each time you run

this, it'll just generate a new random seed.

That's really good advice, David.

That'll save people a ton

of time and I'm sure some

headache trying to figure

that out the first time.

All right, let's run it as is.

All right, so we're looking for some

upbeat indie pop with some warm

female vocals, some bright electric

guitars, punchy drums, melodic bass.

Sounds like it'll be pretty good.

So, a 2-minute, 28-second song

was generated in 59 seconds.

So you can generate music in less

time than the music itself takes.

That's right. So maybe in the future you

can have infinite music.

Let's have a listen to the song.

Not bad.

Pretty good. Now available on Spotify.

Hey Ilan, that's a nice shirt.

It's reminiscent of some good old times.

Yeah, that's right.

Except for that, it's the modern times,

available now at devilwearsproduct.shop.

Nice.

So this works very well with

other genres, of course.

Let's give it a listen.

And one of the weaknesses of this

model and other music generating

models that I've tested, they want

really hard for there to be lyrics.

So even if in the lyrics

portion of the prompt,

you just say instrumental

in square brackets,

it refuses.

It'll generate gibberish,

and it'll start reading part of

the system prompt, like the

actual music style prompt.

Exactly.

That's interesting,

definitely something for

them to fix, because I was

also going to say that where

I can imagine this being

really useful would be lo-fi

hip-hop, which a lot of people

use as concentration music

when they're working,

and it's already a place where

AI is kind of taken over,

because you don't need

much for that to work.

I can imagine having an

endless lo-fi hip-hop

generator that just runs while you work.

So important to know is that this model is

shared under the Creative

Commons license.

And in the details here,

it says that it's fine to use

for non-commercial

sharing and reproduction.

But you cannot use this

for commercial use.

You're not allowed to use this for your

new ad or to generate profit on Spotify.

Good to know.

So there you have it.

Thanks, David. That's how you use YuE2.

Seems like an amazing model,

one that works super well.

We'd love to hear how people

are using it themselves.

Drop us a link.

We'll see you next time.

See you then.