AI Generated Anime Art Is Improving Too Fast [NijiJourney & More]

actual words are overrated it is not

even needed for AI to generate

illustrations anymore what's going to be

next emojis on a more serious note this

is actually possible however it is only

because of the new stable diffusion

model called anything V3 that surface on

the internet a little more than a month

ago its origin is a known so don't ask

where's version 2 or version 1. all we

really know is that it appeared out of

nowhere on a Chinese forum and that it

was most likely fine-tuned on the leaked

novel AI models anything V3 is probably

the best model for generating Anime

styled images and in my humble opinion

the quality of the generated images are

much better than any other models we

have seen at that time closed-sourced or

open source spoiler alert at that time

the reason that I believe anything V3 is

better than any other anime focused

models such as anime 4 latest and waifu

diffusion V 1.3 is not just that it is

slightly better at generating Anatomy it

is also because it handles lighting and

shading way better than what we have

seen 4 especially if you use their vae

checkpoints on top AFL May sometimes

create more stylistic patterns but from

my own experience dealing with AI

generated images being able to produce

good lighting and shading requires more

luck than anything else which usually

involves heavily bracketing detailed

lighting and praying to the almighty air

god that the outcome gives you something

decent maybe people get addicted to

generating air because it feels like

opening a personalized loot box but

anything V3 is like when the loot box

always gives you the legendary tier

items from pix V top monthly level of

lighting to trending on Art station

level of stylization all you need to do

is just to specify it however anything

V3 is only good as a standalone model it

does very bad if you use it for textual

inversion or for fine tuning when

compared to using AFL however what is

very interesting is that when you train

a textual inversion on AFL and then use

it to evaluate it through anything V3 it

tends to generate significantly better

results which is definitely something I

did not expect anything V3 is also very

good at generating Landscapes it has

most likely been fed enough images with

anime backgrounds which meet any

generated image excluding humans feels

like it was taken straight out of a

visual novel anything V3 is very strong

at generating scenery with great

lighting details which makes the

generated images look way more

believable I could definitely see image

generating AI like these being used by

Indie Games visual novels or even some

small anime studios in the near future

because background illustrations are one

of the most tedious parts of making an

anime related stuff what's even better

is that people somehow min max textual

inversion even further for a quick

reminder a few videos ago I introduced

dream booze and in a lot of ways it's

very similar to your typical textual

inversion technique where you embed a

special character and a subject together

dreambooth also find teams of

pre-trained models so that it can be

good at generating a specific subject or

art style while only using less than 10

reference images of the subject during

training in quote-unquote production you

can use easily tell the difference

between them because dream Booth

generates a DOT ckpt model while textual

inversion only generates a DOT PT model

which also needs a DOT ckpt model to run

on to simplify the explanation of dream

Booth even further you just need to give

an AI less than 10 images to learn

sometimes as little as four of a

particular subject or specific art style

they'll be able to recreate it in

different settings easily compared to

before where it took up to around tens

or hundreds of reference images for

textual inversion and maybe even a few

thousands for a proper fine tuning less

than 10 is a huge leap pretty amazing

right while now you only need one after

two months stream Booth was proposed

dream artists came out of nowhere and by

proposing a modified textual inversion

Network it once again stretched the

limit of what AI can do let that sink in

only one image is needed for textual

inversion absolute bunkers the results

that the authors showed were incredibly

impressive my mind was blown out of the

planet when I was in able to reproduce

the same result using the model the

author shared on their GitHub just look

at these they do not look like something

you should be able to achieve using only

a single reference image of course it is

also thanks to training on AFL and

evaluating on anything V3 that this

level of quality was produced but to be

completely honest with you all I still

cannot fathom the fact that it only

takes one reference image to reproduce a

character with such a high level of

quality and coherency here are some

examples to compare the good old textual

inversion and the cool new dream artist

but if you still prefer quality over

everything else other improved textual

inversion or drink Booth methods may

still provide better results as well as

a more accurate representation of your

subject so dream artist is not really a

replacement for the older methods it's

pretty much as a pure Flex there are

still quite a few other downsides to

dream artists though as you might expect

which to be fair is completely

reasonable when you only need a single

image as a reference some of the

downsides include one it has a super

long training time that requires at

least 40 minutes or 2 hours depending on

if you are running the Reconstruction

mode or not two there are crap loads of

parameters you have to get it right

in order to get a decent result or else

your result will look butchered if the

settings were slightly off I actually

attempted to replicate the author's

model and the absolute closest I could

get was only to this level of accuracy

three this one can be a bit

disappointing for the non-webs to learn

but dream artist only works well on

illustrations specifically anime

illustrations as you can see in this

General comparison which Compares

textual inversion and dream boost

provided by the author themselves the

Anime styled R is definitely more on

point as you can see the talk is a lot

less consistent than the anime

characters so you might be thinking why

a lot of anime stuff is closely related

to AI generated art well this may be

thanks to how popular and appreciated 2D

anime related illustrations are in the

online culture as well as that's the

fact that the most well tagged data on

the internet that is open source is

illustrations also it cannot help that

companies cannot generate something that

Strays so far into the gray area of

generating copyrighted content which is

why most of the anime AI models are

released either from unknown sources or

made by someone within the community but

perhaps Journey thinks otherwise

if you haven't heard of it Nietzsche

journey is a collaboration between

mid-journey and spellbrush the Very

spell brush that created waifu Labs

which was the coolest private research

that generated the best anime

illustrations with Gan before diffusion

even popped back into existence I was

lucky enough to be one of the people

invited to test it out during the close

beta and I have gotta say needy Journey

probably contains the best and much more

stylized anime related AI art I have

ever seen on the internet right now it

is very confident in plenty of different

art styles and I am not joking when I

talk about stylization lighting

Reflections Shadows coherency it is

something that I have never ever seen

before and it made anything V3 look like

an amateur it also has a very good bow

and image composition understanding

which is something that you rarely see

in other AI models it can create a super

realistic magazine cover or page and

when I first saw that I had to do a

double take because I thought it was

post-process with Photoshop and

surprisingly its AI models are not

afraid of copyright issues I was shocked

at how easily some anime characters

could be generated while also

maintaining a superb quality that is

very hard to replicate with other AI

models so if you want to use and play

around with it you better get going

before the model gets into legal trouble

or has any restrictions put on it

however since they are completely closed

Source I have absolutely no idea how

they have achieved this which is

unfortunate since it means others can

replicate this phenomenal AI model some

people dislike niji Journeys holy for

the reason that they are closed sourced

and charge people money but what do you

think anyways there is still so much

left I want to cover in the realm of AI

generated art like the new stuff coming

out from Save diffusion but I'll keep

those for the next video or else this

video is never gonna end so subscribe

and stay tuned for that and if this

topic interests you and motivates you to

start learning AI or machine learning in

general today's sponsor brilliant is

actually one of the best places to get

started brilliant is an online learning

platform that is basically when your

textbooks become alive it provides a way

for you to learn interactively with

brilliant fun Hands-On lessons in math

science and computer science research

has shown that Interactive Learning

helps you learn six times more

efficiently than watching lecture videos

and I totally agree with that

interactive lessons only help you

visualize problems much quicker but also

are able to illustrate very difficult

concepts for you to comprehend faster

which is what plain textbook or YouTube

videos cannot help you with back in my

high school days I actually used

brilliant along with my Calculus class

because it was much easier to understand

what calculus is on about while you are

being freshly introduced to a new field

of maths brilliant not only have very

helpful diagrams for you but also

interactive elements that actually help

my understanding much faster than

learning the fundamental theorem of

calculus through a wordy mathematical

definition inside a boring textbook not

only that brilliant also provides a

clear road map on different subjects

where our knowledge levels from basic

algebra to Advanced multivariable

calculus from programming with python to

artificial neural networks brilliant is

full of stem classes that are usually a

pain to study in but made into a much

friendlier and digestible format so yeah

you can quickly get started on brilliant

by heading to brilliant.org by Cloud to

get started for free with brilliant ever

expanding interactive lessons and to

also support this channel the first 200

of y'all will also get 20 off an annual

membership thank you so much for

watching a huge shout out to Andrew

lesscelius Chris LeDoux

sean7134 and many others that support me

through patreon or YouTube If you have

any questions feel free to join my

Discord and ask there follow my Twitter

if you haven't and I'll see you in the

next one