Is Kandinsky-2 is better than Stable Diffusion?

Summary

В этом видео обзоре рассматривается инструмент генерации изображений Kandinsky 2, который может создавать разнообразные типы изображений, такие как текст в изображение, изображение в изображение, и т.д. Хотя Kandinsky 2 не имеет всех функций, доступных в стабильной диффузии, он предлагает различные стили и возможности, которые могут заинтересовать пользователей. В видео также сравниваются результаты работы Kandinsky 2 с другими популярными моделями, такими как стабильная диффузия, показывая различия в стилях и качестве изображений. В конце авторы предлагают использовать Google Colab для работы с Kandinsky 2 без необходимости установки программного обеспечения на локальный компьютер.

Takeaways

• 🖼️ Kandinsky 2 - это инструмент генерации изображений, который может создавать различные стили изображений, включая текст в изображение, изображение в изображение, микширование изображений и другие.
• 🇷🇺 Название Kandinsky 2 указывает на российские корни проекта, что может вызвать определенные озабоченности у пользователей по поводу политического контекста.
• 📜 Лицензия Apache 2, которая используется у Kandinsky 2, является довольно разрешительной и позволяет свободное использование и распространение.
• 📚 В репозитории есть несколько примеров ноутбуков Jupyter, которые демонстрируют, как использовать различные функции Kandinsky 2.
• 💽 Для работы с Kandinsky 2 рекомендуется иметь видеокарту с VRAM не менее 10 ГБ для обработки изображений размером 512x512 пикселя.
• ⚙️ Для локальной установки Kandinsky 2 предпочтение отдается использованию Linux с современной видеокартой Nvidia.
• 🔧 Если у вас нет необходимой аппаратной платформы или вы не хотите устанавливать что-либо локально, можно использовать Google Colab для удаленного доступа к среде Linux с поддержкой Nvidia.
• 📁 Установка Kandinsky 2 может потребовать создания отдельной среды Python с использованием таких инструментов, как Anaconda, Miniconda или venv.
• 🔄 Kandinsky 2 предлагает функцию микширования изображений, что позволяет комбинировать различные изображения в одном изображении.
• 🎨 В интерфейсе Kandinsky 2 есть возможность в реальном времени управлять параметрами генерации изображений с помощью слайдеров и текстовых полей.
• 📈 В сравнении с другими моделями, такими как стабильный диффюзный (Stable Diffusion), Kandinsky 2 показывает хорошие результаты при генерации сложных изображений с учетом стиля и деталей.

Q & A

Kandinsky 2 является каким типом искусства искусственного интеллекта?

Kandinsky 2 относится к типу искусства искусственного интеллекта, который фокусируется на генерации изображений.

Какой тип лицензии имеет Kandinsky 2?

Kandinsky 2 имеет лицензию Apache 2, что является довольно разрешительной.

Какие функции предлагает Kandinsky 2?

Kandinsky 2 предлагает функции, такие как преобразование текста в изображение, изображение в изображение, смешивание изображений, заполнение участков изображения и вставка текста в изображение.

Какой объем видеопамяти (VRAM) рекомендуется для работы с Kandinsky 2?

Для работы с Kandinsky 2 рекомендуется иметь видеопамять объемом не менее 10 гигабайт.

Какой операционной системы и оборудования предпочтительнее использовать с Kandinsky 2?

Для лучшей производительности, удобства и совместимости с Kandinsky 2 лучше всего использовать Linux с современной видеокартой Nvidia.

Как можно использовать Kandinsky 2 без установки на локальном компьютере?

Для использования Kandinsky 2 без локальной установки можно использовать Google Colab, который предоставляет удаленный доступ к комбинации Linux и Nvidia бесплатно.

Какой командой можно установить Kandinsky 2 на локальном компьютере?

Для установки Kandinsky 2 на локальном компьютере можно использовать команду `pip install`, указанную на странице GitHub проекта.

Какие дополнительные библиотеки могут потребоваться для запуска всех ноутбуков Kandinsky 2?

Для запуска всех ноутбуков Kandinsky 2 может потребоваться установка дополнительных библиотек, таких как Jupyter Notebook, ipywidgets, open CV и protobuf.

Какие примеры использования представлены в ноутбуках Kandinsky 2?

Ноутбуки Kandinsky 2 содержат примеры использования различных функций, таких как генерация изображений из текста, преобразование одного изображения в другое, заполнение участков изображения и другие.

Какие сравнительные тесты проведены между Kandinsky 2 и стабильной диффузией?

В скрипте проведены сравнительные тесты по генерации изображений с использованием различных стилей и тем, таких как киберпанк, акварель, сурреализм, реализм и другие, для определения эффективности каждого из методов.

Какие результаты показала сравнительная работа Kandinsky 2 и стабильной диффузии?

В результате сравнения, Kandinsky 2 показал лучшие результаты во многих случаях, особенно в стилях, требующих деталей и специфического восприятия, таких как киберпанк, сурреализм и реализм.

Outlines

00:00

🌟 Введение в Kandinsky 2 и его возможности

В этом абзаце представлен обзор проекта Kandinsky 2, связанного с генерацией изображений. Рассматриваются различные функции, такие как преобразование текста в изображение, смешивание изображений, заполнение масок и другие. Также упоминается, что репозиторий имеет Apache 2 лицензию и предлагает две версии моделей: 2.0 и 2.1. Обсуждаются системные требования, включая необходимость достаточной видеопамяти для работы с изображениями размером 512x512. Рекомендуется использование Linux и современного Nvidia GPU для лучшей производительности. Кроме того, упоминается возможность использования Google Colab для удаленного доступа к системе Linux и мощности Nvidia.

05:02

📚 Установка и настройка среды Kandinsky 2

Абзац описывает процесс установки среды Kandinsky 2, включая создание нового окружения с использованием Anaconda, установку основной библиотеки и дополнительных компонентов, таких как Jupyter Notebook, ipywidgets и OpenCV. Также упоминается необходимость установки версии Protobuf 3.20.0. Далее автор копирует примеры ноутбуков из репозитория и обсуждает организацию файлов и каталогов для удобства работы.

10:04

🖼️ Испытание функций Kandinsky 2 с примерами

В данном абзаце автор демонстрирует, как использовать различные функции Kandinsky 2, включая преобразование изображений, заполнение масок и другие. Показывает результаты работы с ноутбуками версий 2.0 и 2.1, обсуждает интерфейс и возможности каждого из них. Также проводится сравнение с другими инструментами, такими как стабильное распространение, и автор дает своё мнение о качестве полученных изображений.

15:05

🎨 Тестирование стилей и сравнение с Stable Diffusion

Автор проводит серию тестов, сравнивая результаты работы Kandinsky 2 с Stable Diffusion на разных стилях и(prompt) запросах. Оценивается качество изображений, созданных с использованием различных стилей, таких как киберпанк, реалистичный, акварель, сюрреалистичный и другие. В ходе сравнения автор дает предпочтение результатам, полученным с использованием Kandinsky 2 в большинстве случаев.

20:06

🎭 Попытка создания антропоморфных и стилизованных изображений

В этом разделе автор пытается создать более сложные и стилизованные изображения, такие как антропоморфный счастливый огонь, римский солдат и другие. Проводится сравнение с Stable Diffusion, и автор оценивает, как хорошо каждая из программ обрабатывает сложные запросы и создает изображения в нужном стиле. Заключение о том, что Kandinsky 2 более удачно справляется с созданием стилизованных изображений.

25:06

🥪 Тестирование функций микширования изображений

Автор рассматривает функцию микширования изображений Kandinsky 2, объединяя римского солдата с ланцем и бутербродом. Показывает, как программа комбинирует элементы изображений и создает новый образ, который является миксом оригинальных элементов. Хотя автор не особо доволен окончательным результатом микширования, он подчеркивает потенциал данной функции для других пользователей.

Keywords

💡AI

Искусственный интеллект (AI) - это область информатики, которая разрабатывает программное обеспечение и машинное оборудование, способные функционировать на уровне человеческого интеллекта или лучше. В видео речь идёт о прогрессе в области AI, особенно в создании больших языковых моделей и генерации изображений, что демонстрирует разнообразие и развитие технологий.

💡Изображение

В контексте видео, изображение - это результат работы алгоритмов генерации, таких как Kandinsky 2, который превращает текстовые описания или другие изображения в визуальное представление. Генерация изображений является ключевым элементом в демонстрации возможностей AI, как в случае с Kandinsky 2.

💡Kandinsky 2

Kandinsky 2 - это программное обеспечение для генерации изображений, упомянутое в видео. Оно способо создавать различные стили изображений на основе текстовых описаний или других изображений, что показывает современные возможности искусственного интеллекта в области компьютерного зрения.

💡Текст в изображение

Текст в изображение (text to image) - это технология, которая позволяет на основе текстового описания создавать визуальное изображение. В видео это один из основных функционалов Kandinsky 2, и он демонстрирует, как AI может интерпретировать текст и создавать соответствующие визуальные сцены.

💡Изображение в изображение

Изображение в изображение (image to image) - это процесс, при котором исходное изображение преобразуется или модифицируется с использованием AI в соответствии с текстовыми prompt или другими образами. В видео это один из функционалов Kandinsky 2, используемый для демонстрации возможностей преобразования и слияния визуальных данных.

💡Слияние изображений

Слияние изображений (image mixing) - технология, которая объединяет два или более изображения в одно, часто с целью создания нового визуального содержания. В контексте видео, Kandinsky 2 может смешивать изображения, создавая уникальные визуальные композиции.

💡Заполнение изображения

Заполнение изображения (in-painting) - это процесс редактирования изображений, при котором AI заполняет или восстанавливает части изображения, которые были удалены или повреждены. В видео это функция Kandinsky 2, используемая для демонстрации восстановления и улучшения изображений.

💡Лицензия Apache 2

Лицензия Apache 2 - это одна из самых популярных открытых лицензий на программное обеспечение, которая предоставляет большую степень свободы использования, распространения и изменения программного кода. В видео упоминают, что Kandinsky 2 имеет такую лицензию, что означает открытость и доступность его использования и расширения функциональных возможностей.

💡VRAM

VRAM (видеооперативная память) - это тип оперативной памяти, используемый графическими процессорами (GPU) для хранения и обработки графической информации. В контексте видео, VRAM является важным показателем для работы с программами по генерации изображений, так как большие объемы VRAM позволяют обрабатывать более сложные и высокие разрешения изображений.

💡Google Colab

Google Colab - это облачная платформа от Google, которая предоставляет доступ к ресурсам для машинного обучения и анализа данных без необходимости установки программного обеспечения на локальный компьютер. В видео упоминается как альтернатива локальной установке программного обеспечения для работы с Kandinsky 2, позволяя пользователям использовать его в браузере и получать вычислительные ресурсы Google.

💡Анаконда

Анаконда (Anaconda) - это дистрибутив Python, который включает в себя среду выполнения Python, менеджер пакетов и многочисленные научные библиотеки. В видео автор использует Anaconda для создания изолированной среды разработки, в которой устанавливается и запускается Kandinsky 2, что обеспечивает удобство и изоляцию проектов.

Highlights

Kandinsky 2是一个图像生成工具，可以产生相当合理的图像，尽管它没有像Stable Diffusion那样的所有特性。

Kandinsky 2具有Apache 2许可，这是一个相当宽松的许可，并且拥有许多特性。

该工具支持文本到图像、图像到图像、图像混合、图像修复等功能。

Kandinsky 2的文本到图像功能与Dali 2和潜在扩散的混合体相似，是最受欢迎的特性之一。

存在两个模型集合，一个用于版本2，另一个用于版本2.1。

版本2.1比2.0稍微轻量级，对于512x512图像，它使用了8.7GB的VRAM。

对于最佳性能和兼容性，建议使用Linux加上现代Nvidia GPU。

Google Colab提供了一个免费的远程使用Linux加上Nvidia GPU的选项。

如果不想本地安装或没有所需硬件，Google Colab是一个很好的选择。

使用Anaconda可以轻松管理Python应用程序的单独环境。

Kandinsky 2提供了Jupyter笔记本示例，方便用户开始创建自己的程序。

版本2.1提供了一个稍微不同的界面，使用Jupyter Lab而不是Jupyter Notebook。

Kandinsky 2的图像后处理功能允许用户在保存图像之前进行调整。

Kandinsky 2在处理复杂提示时表现出色，如生成一个穿着格子裙和黑色军靴的大理石雕像老鼠。

在与Stable Diffusion的比较中，Kandinsky 2在多个风格和主题的图像生成中表现更好。

Kandinsky 2的图像混合器允许用户混合两个或更多的图像，创造出新的视觉效果。

尽管Kandinsky 2在某些混合图像的输出中可能不是每个人的口味，但它提供了强大的图像生成和编辑功能。

Transcripts

hello and welcome to More nerdy rodent

geekery you might have noticed that

loads of AI stuff seems to be released

almost every day now with lots of focus

around large language models and image

generation

Kandinsky 2 falls into the image

generation camp and can produce some

fairly reasonable images sure it hasn't

got all the features of something like

automatic 1111 for stable diffusion but

sometimes it's just nice to try other

stuff isn't it I mean variety is the

spice of life and all that as you may be

able to tell from the name this is from

Russia so you may not want to go

anywhere near it as a supporter of

Ukraine I know I was in two minds about

even covering this at all but at the end

of the day this isn't a politics Channel

and we're all just interested in the

nerdy stuff right okay the main features

include an Apache 2 license which is

fairly permissive and it's got a whole

bunch of features too if we scroll down

here we can see some things we've got

text to image we also have image to

image there's some image mixing or image

fuse as it says there and also some

image in painting as well here we scroll

down we've got some texture image we've

got an in-painting example there they've

got the mask at the bottom and it's

filled that in underneath like it says

on the GitHub page here this is sort of

a mix between Dali 2 and latent

diffusion I really did get some Dali 2

Vibes when using text to image and that

is the feature I like the most out of

this repository the other features I

personally classify as a bit meh but

that is just my personal opinion you may

love image mixing and this is indeed a

pretty good implementation of it there

are actually two sets of models one for

version two and a another set for

version 2.1 and if you do download

everything that totals around 17 gig

personally I would ignore 2.0 and just

go with 2.1 vram wise you'll want

something at least capable of playing

video games from 2023 so ideally 10

gigabytes of vram or more 2.1 is

actually slightly lighter than 2.0 and

for a 512 by 512 image it used 8.7

gigabytes of vram for me I have no idea

if this will work on a Mac but it does

seem unlikely I also don't have an AMD

GPU so again I can't say if it will work

with one of those either but maybe if

you installed the correct version of

pytorch afterwards who knows do let me

know as with practically everything AI

these days your very best option for

performance ease and compatibility is to

use Linux along with a modern Nvidia GPU

other operating system options good and

include windows subsystem for Linux also

known as WSL or if you really love

headaches then you could use a native

Microsoft Windows installation on the

other hand if you don't want to install

anything locally or you don't have the

required hardware then Google colab may

be the thing for you there's a link

there you click that open and there you

get your very own Linux plus Nvidia

power combo to use remotely for free and

it's very very easy to use as you can

see there's just a few cells so you

click play on the first one click play

on the second one click play on the

third one and there is where you type

your prompt so in this example Redcat 4K

photo

768x768 and there you get rather a nice

picture of a red cat

as another option there is also a link

there to demo which will take you over

to this Fusion brain demo so that's

another way that you can start playing

with this model without having to

install anything if like me you are

installing this python app locally then

all you really need to do is run that

one command there on your computer and

you've got it done however I do like to

keep all of my python apps in separate

environments and thanks to Anaconda that

is made really really easy if you like

separate environments too but don't want

to use Anaconda then mini conda or VNV

are great options too basically whatever

you feel the most comfortable with using

is what you should use we're all about

Freedom of Choice here in my case I'm

using Anaconda so all I need to do is

open my anaconda prompt create a new

Anaconda environment and then I can run

that single pip install command in my

managed environment the first command is

Condor create I've called mine Kandinsky

too because I'm great at names and I of

course have already created it so I'm

just going to say no there but you would

say yes the next thing to do is activate

that environment there it is conda

activate Kandinsky 2 and it has changed

from bass to my new Kandinsky 2

environment now I've got a nice new

python environment to install things

into let's actually install it there it

is that one pip install command from the

GitHub page just copied and pasted that

from there and we'll let that run for a

few seconds

okay there you go you've got the main

thing installed however this does come

with a load of Jupiter notebooks as you

can see there so what we're going to do

as well as that is also install Jupiter

notebook there I've got Condor install

minus C Condor Forge notebook ipy

widgets and also open CV again that will

just take a few seconds to install

final thing I found it needed on one of

the notebooks I can't actually remember

which one it was now was Proto buff as

well so a quick pip install of protobuf

version

3.20.0 and that will give you everything

that you need to run all these notebooks

right if we scroll down a little bit

here we've got the how to use section

like it says there check out our Jupiter

notebooks with examples in the notebooks

directory it also has some examples

there that you can just copy and paste

and start creating your own programs

coding up stuff in your own environment

to match whatever your personal

requirements are I on the other hand am

super lazy so instead I just started by

cloning this repo and then making a copy

of that notebooks directory on my

computer I keep files organized in a

directory structure that I find easy to

use but this may not be a structure that

you want to use so don't worry you don't

need to copy it especially if your name

isn't nerdy on your computer you can use

whatever directory structure or lack

thereof is most suited to you I know

it's that whole Freedom of Choice thing

again you do you Boo for me I keep my

GitHub stuff here in this directory that

I created myself and called GitHub it's

also where I'm going to run the get

command from and I'm pretty sure get

comes installed as standard with pretty

much every Linux distro nowadays but if

for whatever reason you don't have git

installed then you can just run conda

install git now and that will do it for

you next up is to run that git clone

command and that will download all the

files from the repository I of course

have already done that so I won't be

running it now

and will instead move straight on to the

next step which is changing directory

into that newly created area so there I

am in GitHub Kandinsky too now I've

already copied it let's just show you

the contents there I've already copied

the notebooks directory to my notebooks

but you can do the same if you want to

now I'm going to say upfront I found

notebooks a little bit janky and I'm

going to show you the ones for 2.0 quite

quickly that way you can decide for

yourself if you do indeed want to use

2.0 at all I find these notebooks a

little bit janky so I'm just going to go

through them really quickly and only for

2.0 this way you can see exactly how

they work and also decide for yourself

if you want to use 2.0 at all

right so first up I'm going to change

into my notebooks directory which is

just a copy of notebooks I'll show you

the contents there I've just got all

those notebooks and also a couple of

example images to run those notebooks

I'm just going to run the Jupiter

command with the notebook option and a

couple of seconds and there it is it'll

start up automatically in my web browser

and show me all those notebooks there

we'll take a look at the image 2 image

one first

one thing to note is that the default

cache directory is set to slash TMP this

is great however it does mean that any

files you download will get wiped out

when your temp directory gets cleaned

thus what I like to do is change the

cache there in all of these examples to

a permanent storage location so you'll

see that here I've got the cache

directory and by default it isn't even

there but like I say it's set to slash

temp so there I have home nerdy GitHub

Kandinsky 2 cache and that I have

changed in all of these notebooks as

with Google collab you just have a

number of cells and you click run but

what I can do here is just do sell and

then run all and that will run all these

cells in order as you can see there we

should have a photo of a bearded man

come out at the end now this is doing

image to image and it is image opening

that image.png so let's show you that

one that is there so that's the image to

image it's doing you might be familiar

with image to image if you've used

stable diffusion at all but it's

basically using that image as an input

and also merging it with the text prompt

all right let's have a look see what it

has produced loads and loads of text

it's fine you can ignore all that

because down at the bottom there we've

got a photo of a bearded man so there

it's done the strength 0.85 so it's only

changed it a little bit number of steps

50. this is all very familiar stuff if

you have indeed used stable diffusion so

yeah image is the one got their photo of

a bearded man like I say not my favorite

function this nor is anything at 2.0 but

there it is working

okay moving swiftly on the next one is

the in painting notebook again here I'm

just gonna run all the cells and

obviously I've updated that cache

directory as well in this case instead

it's done a bit of in painting so

there's the mask it's basically the

bottom bit of that and I've impainted

that with a bronze statue as the

original image so it's cut that off

change that made him into a bronze

statue not too bad I do quite like the

in painting a little bit but still not

quite as good as stable diffusion okay

so that's 2.0 out of the way those are

the two things you can do with those

let's take a look at 2.1 now but with a

slightly different interface instead you

can of course install this in exactly

the same environment so here I'm also in

Kandinsky too but I'm going to git clone

this repo there as you can see from my

directory listing I have of course

already run that command so I can just

move straight onto the next step which

is changing directory into there there I

am change directory into the advanced

notebooks now you want to install a few

extra things here so once again you will

run pip install minus r requirements.txt

and you'll have to wait a few seconds

for that to run through

at the end of that certainly for me

anyway I also install the font Roboto

pip package too because it uses that

font later on now this uses Jupiter as

well but rather than Jupiter notebook

there we are going to run Jupiter lab

instead

as I did with the example Jupiter

notebooks I've also changed my cache

directory in here in this case it's in

main.py so if you double click that

you'll get that file open and there I

have set my cache directory

now this is an optional step but if like

me you get errors with the font if you

go into the UI directory then it edits

the init.py file there it is that's this

one here and then you scroll down to

where it says Roboto regular and there

as you can see on the image font line I

have instead given the absolute path to

that Roboto regular ttf font file

another thing you may like to do is like

me have this dark theme so up here at

the top settings theme it's Jupiter lab

light by default which absolutely blinds

me I can't see using that at all so I

use Jupiter lab dark

if you do want to get that nice

real-time updating GPU widget then you

can uncomment that line and install that

but personally I've already got one on

my screen so I don't need to just run

that cell and then eventually the user

interface will appear now do note this

will download a whole bunch of models to

start with if you haven't so the first

time this will take a couple of minutes

before the interface appears

and there it is as you can see now we've

got a whole bunch of Sliders and things

that you can have menus with and you can

make the prompts bigger so it's an ever

so slightly better interface for playing

with this you've also got this image

post processing at the Top If you expand

that and then expand image saving if you

want to you can save out all of your

images now one problem with this however

is that the directory has to exist

already it won't create it for you so do

make sure that you create that directory

first before specifying it in save path

all right so I'm just going to shrink

that down for now because I'm absolutely

fine with those we'll take off the save

because I'm not going to save these and

then we've just got the normal interface

right I've also got stable diffusion

running here as well so I can do a

little comparison of this versus stable

diffusion 1.5 and there we are hello

good old stable diffusion alright so

I'll be switching back and forth between

these two tabs let's start off with

something fairly easy make this prompt

box a little bit bigger and there I've

got a cyberpunk rodent is wearing a top

hat whilst dancing on an oak table

inside a South London pub in summer if I

do a generate on that one and also on

stable diffusion at exactly the same

time as well we can see which produces

the best image as you can see on this

one by default it does create six images

so we'll have a few to look at and there

we have our wow that's actually really

quite good obviously the hands are

absolutely crazy but that really is a

cyberpunk rodent wearing a top hat

dancing on an oak table inside a South

London pub that's cool let's have a look

at some of the others there he is he's

on the table there's some more tables in

the background so that they're all quite

similar they're a little bit different

but they all they all have a certain

sort of feel to them they all they're

all the same but he's dancing on the

table all right let's have a look at

stable diffusion and their stable

diffusion and

hmm

I don't think that's quite as good

to be honest

I don't think that's quite as good

because in all of these the rodent is

actually dancing on the table he's got

his steampunk hat going on this one he's

got the top hat and of course you know

the hands are a little bit manky but

he's not he's not really on a table all

right let's be fair to stable diffusion

and we'll give it another couple of goes

there maybe I should increase this batch

count slightly let's let's pop that up

to four and uh but once again you see

it's not on the table it's not on the

table so I'm going to give that one to

Kandinsky purely because the rodent is

on the table and that's what I asked for

all right let's try again with another

prompt let's put this one in this is

this is also going to be a fun one so

here we have a professional photo of a

kitten painting a watercolor image of an

open cat food tin on a white canvas

supported by an old wooden easel and

again I've popped the same thing into

stable diffusion as well and see how

both of these do let's take take a look

at this first one in Kandinsky there it

is it's it's okay it's not bad I mean

the kitten isn't painting a watercolor

image of the cat food tin but it is a

painting it's got the canvas it's got

the easel it's got the tin not too bad

there the kitten itself seems to be the

painting so no again no we've got a

painting rather than cat food in and

some weird paint brushes at the top I

mean they're quite good they are quite

good that that one's not bad it's close

it's close it's close all right it's

they're quite good images but let's have

a look at stable diffusion all right

we've got four there and oh

oh I think that one is probably the

closest I mean the kitten isn't actually

painting but it does have kind of a cat

food tin that's kind of got a cat food

tin as well and that's just a human

painting a kitten I think I'm gonna give

that one to stable diffusion personally

although I do quite like some of the

textures on that let's scroll up again

and try another one this time I've got a

moose head emerging from the birthday

cake on a metal table in a 1960s style

American kitchen okay let's have a look

see how Kandinsky did

all right that's that's nice and weird

we seem to have a bit of a party moose

going with a sort of party hat there

it's not quite emerging from the cake

but it still looks pretty cool oh not

that one yeah that's a moose cake that's

a moose cake I quite like the the sort

of Shades and the some of the lighting

on that as well it's it's got a certain

sort of look to it that's that's kind of

nice

yeah okay anyway

um yeah we've got got more moose there I

think I think that's done quite well I

think that's done quite well all right

let's have a look at stable diffusion

here is stable diffusion oh I don't know

I don't know that one's quite good but

it's made the mousse and it's and that

one's quite realistic and it's got the

moves cut that one's pretty good

that one is pretty good but it doesn't

have that 1960 style American Kitchen

theme to it that's that's more modern

and ah I'm I'm actually going to give

that to Kandinsky I think because I like

the 1960s style theme it's given to it

it's got the mousse I think I just

prefer the sort of color and the

lighting in that it's it's more it's

more like what I was thinking about in

the in the 1960s Style

let's try another one to see how well it

handles complex prompts we'll pop this

one in this time we have a full color

highly detailed high resolution

professional photo of a marble statue

depicting a cute mouse wearing a red and

white gingham dress and massive dark

polished black leather army boots

let's take a look at this do we have any

marble statues wearing red and white

gingham dresses and black army boots uh

yes yes we do that is pretty much

exactly what I asked for all right well

done kamdinsky I don't think you can do

any better than that that's pretty good

okay the boots aren't quite so black

there but they are there and I think

those are amazing it's going to be

difficult for stable diffusion to beat

that all right let's have a look at

stable diffusion

oh it's it's close it's close but these

are more like stuffed toys and there

we've got two I only wanted one it's got

the red and white gingham dress and the

black army boots but that is not a

marble statue I'm afraid I'm really

sorry stable diffusion but I'm gonna

have to give that to Kandinsky once

again let's drop the batch count a

little bit here so we don't have to make

so many and we'll try doing a few

different styles okay so here I'm just

going to do a portrait of a woman in the

style of a particular a painter let's

see how well it has handled that style

oh that's that's quite nice I quite like

that that's it's got quite a sort of

painted texture fit it's definitely the

style it's got the same colors

those are pretty good I do quite like

those I quite like those all right let's

have a look at stable diffusion oh also

done quite well whoops also done quite

well but it's

I don't know that's tricky actually

that is very tricky that's a I think

that's a better representation of the

style but that has more of a sort of

textured

feel to it so I don't know I don't know

that's tricky that is very tricky I like

both of those equally I think they've

both done very well

that's captured more of the style

perhaps accurately and that's done more

of the painting so I'm just going to

call that even between the two now

that's a fairly easy style so let's try

something slightly more difficult now I

know I often have issues generating

watercolor Styles so let's try a

watercolor art style painting of a moldy

sandwich on a park bench let's have a

look-see here oh

that's not too bad actually that's not

too bad it's got the sort of watercolor

blotchiness to it oh

I do quite like those that's that's the

sandwich is the bench so it's it's a

little bit uh it's a little bit off on

the subject but as far as the style that

one's closed that's got a sandwich on

the bench as far as the style I think

that's done very well let's have a look

at stable diffusion audit I can already

see that's not quite as good all right

uh no no is the answer to that I don't

even really need to look any further

that that's not watercolor but it's

partly down to the model obviously but

uh yes Kandinsky wins clearly there all

right let's try another style we'll do

that same thing again we'll do the

sandwich on the bench but this time in a

surreal art style now how surreal is

this gonna be oh that's quite nice

that's quite nice I I do quite like the

colors in Kandinsky there's just

something about the the slightly

different feeling it's got to that's

perfect that's a yeah that's that's a

sandwich on a bench that's the wheel

okay so it seems like surreal also

interprets as dripping but I like that

one I like that one quite a lot all

right let's have a look at stable

diffusion is it um no sorry stable

diffusion you're once again

disappointing I mean it's nice it's got

the sandwich on the bench but it's not I

don't really cast that as surreal no I'm

going to give that style to Kandinsky

once again

let's let's change this up ever so

slightly once more we're gonna do a

realistic graffiti art style painting of

a moldy sandwich but this time it's

going to be on a park bench in New York

at midnight

how well has it done this time for our

graffiti art style painting of a moldy

sandwich on a park bench in New York at

midnight

um

actually slightly stunned by that

because that's brilliant

that is really cool

all right that's

that one's not quite as good that's not

quite as good but the style

the style and the feeling of it I love

those that's fantastic

that is fantastic it's that little

drippy bits down there oh that's nice

that is a really nice style all right

let's have a look at stable Fusion

okay all right

um yeah not as impressed sorry I'm gonna

give that realistic graffiti art style

at night see these stable diffusion ones

aren't even at night that's that's

daytime that's that's completely

Kandinsky

let's let's go for something really

really difficult now then this this

comes out quite interestingly so animate

cartoon style illustration of some

anthropomorphic happy fire with cell

shading and detailed accurate nerdy and

handsome and high quality and amazing

which is of course trending on GitHub so

I'm hoping for a bit of an anime cartoon

style theme to this let's see how these

two compare

anthropomorphic fire how does it oh

that's quite nice that's quite nice it's

it's some nerdy anthropomorphic fire

yeah that's nice I quite like it got a

little bit of a grin going on there some

some cheeky nerd fire yeah I like like

The Cheeky nerdfight all right let's

have a look at stable diffusion oh I'm

gonna be disappointed again aren't I

it's given me loads and loads of faces

and then I mean they're kind of fiery a

little bit but

um that's not really what I wanted okay

I'm sorry I'm gonna give that to

Kandinsky once again okay so we've done

loads of Styles now we've done

watercolor and impressionist and surreal

and realistic and anime let's try

something slightly realistic how can we

go on the realistic here I've got a high

quality high resolution details

professional digital photo of a Roman

soldier he's middle-aged he has dark

hair and there is also a spear let's

generate some of those and see what

happens so the realism test how good is

it it's not bad the eyes are a bit wonky

but then eyes are a bit wonky in stable

diffusion as well it's it's not too bad

it got you know some some battle stuff

going on there that's that's not too bad

it's okay it's okay all right let's have

a look at stable diffusion oh we haven't

got faces there we've got actual

characters so let's let's see there

um

hmm

that's tricky

I mean obviously they're a little bit

wonky uh that's quite realistic that's

quite I mean I prefer those again for

the style for the for the aesthetic of

the output I prefer Kandinsky the image

mixer much the same thing again we'll

just run this cell to have the interface

appear after a few seconds okay there it

is for the image mixer again you've got

the same options if you want to do the

saving and including generation text at

the bottom let's just close that down

again so this is mixing images so you

can mix one or more images there you can

upload one upload another one and keep

adding images into the mix by default it

has two so let's just use these defaults

and we'll select our little soldier guy

got a soldier guy there and then let's

also select one of those delicious

sandwiches so we'll select that upload

there pick a delicious sandwich so now

we're doing a mix of our Roman soldier

with a spear and one of the sandwich is

mix two images what's it going to do I

don't know let's find out

and there you have it a mix between a

Roman soldier and a sandwich it's sort

of taken the the bench turned that into

a chair although that's a bench as well

and then it's got the sandwich so it's

done quite a good job at mixing but as

with many of these mixing programs I

don't really like the final output

however one thing you may like is this

next video