Manual: Creation of artworks using Leonardo.

ai
Stênio Nobre and Communication 2023.1B

1. Objective
This report has the primary objective of presenting a
new tool that has been showing itself to be very competent and useful in
last months, known as Leonardo. This is an artificial intelligence.
focused on creating art and graphic design, a new type of tool that
has been developed recently that generates images from
command prompts.
The use of this in communication can be beneficial in numerous
senses, but this report will mainly address the creation of
image vectors using AI.

2. What are Artificial Intelligences?

Artificial intelligences (AI) simulate reasoning and learning.
humans to perform autonomous and efficient tasks. They can be used
in complex areas that require advanced data processing and
decision making. Pattern recognition, processing of
natural language, autonomous vehicle driving, and medical diagnosis
These are some of the tasks that AI can perform. Through algorithms.
and machine learning models, AI analyzes information, learns
with past examples and adapt their behavior to improve the
performance over time. Its use is significantly beneficial, but
it also raises ethical and security issues.
AI can be classified into rule-based AI, which
they follow predefined instructions, and AI based on learning from
machines that learn from data. They can be narrow AI, specialized
in a specific task, or general AI, with skills comparable to
humans in various areas. AI already has an impact on our daily lives, such as
virtual assistants and personalized recommendation systems.

2.1. Artificial Intelligences for image creation

Artificial intelligence applied to image creation is a field
in rapid growth. It combines machine learning algorithms and
image processing to generate automated visual content. This
allows AIs to learn to recognize and reproduce visual elements,

1
such as shapes, colors, textures, and artistic styles. A notable example is the
use of convolutional neural networks to generate realistic images. They are
trained with datasets containing real examples and can create
new images in the same style. The DeepArt project is a practical example
thereby transforming photographs into works of art in the style of painters
celebrities.
In addition, artificial intelligence is used to create images
synthetics, such as landscapes, characters, and objects that do not exist in
reality. These images are used in entertainment, games
digital and scientific simulations. Generative adversarial networks (GANs)
are used in this context. They consist of two neural networks that
compete with each other: one generates synthetic images and the other evaluates them
authenticity. With proper training, GANs are capable of generating
highly realistic and convincing images.
Midjourney is an impressive example of artificial intelligence.
focused on generating realistic images from textual descriptions.
Using advanced deep learning techniques, this platform allows
for users to provide a description and, in response, Midjourney
generate a corresponding image. This versatile tool can be
applied in various fields, such as computer-generated art and creation of
characters for games, offering an innovative way to bring to life
visual ideas. Another notable example is BlueWillow, a tool for
image editing powered by artificial intelligence, designed to
simplify the work of designers and editors. This platform offers
automated resources, such as color adjustment, object removal
unwanted effects and application of effects. In addition, BlueWillow allows for
transformation of sketches into realistic images through the use of networks
neural, providing a quick and efficient way to create illustrations
detailed from simple drawings.

3. What is [Link]?

[Link] is a new artificial intelligence that has emerged.

firstly with the idea of assisting game developers
ease in asset creation of high quality, this offers the
capacity for specialization in various topics, allowing users
choose the model that best suits your needs. It is possible
Customize the result by selecting the desired model and specifying
the priority of the request. A unique advantage of [Link] is the

2
possibility to train custom image models, something that
other similar tools do not offer. Additionally, it is possible
share the models with the community, making the tool
collaborative. Although Midjourney is currently a leader in quality of
image, [Link] represents a significant advance in generation of
computer images and is especially useful for designers who
they wish to have complete control over their exclusive models. However, it is
It is important to remember that the final use of each tool is determined
by the user.

4. Registration on the site

To use the AI, you must first register on the site for
gain early access, remembering that this tool is in phase
Beta, it is necessary to enter a waiting list to access it. To
to register, one must access the official websiteLeonardo, and insert your
email you wish to be linked and wait, the most interesting thing would be to insert
an email that you have linked to Discord, for reasons that will be
most commented ahead.
Figure 1: Registration area on the website

Source: Screenshot generated in [Link]

When registering for early access, there will be a period of

analysis of your email, due to a priority order of access. You
you will receive an email in your inbox that will contain some
information about the tool and the community and also a link that
redirects to the AI's Discord server, which can be used to
gain priority in early access.

3
[Link] Server
Upon entering the server, you will have access to the community, which will feature
with support and a dedicated channel for speakers of the Portuguese language, which
can be used to ask for tips, interact, share, and view images
that others shared, etc.
Figure 2: [Link] server

Source: Screenshot generated on Discord

To gain priority access, you must access the channel 'Guide of'
"Server" in the upper right corner. Upon accessing it, there will be direct access to
tutorial on how to obtain this priority.
Figure 3: Priority access tutorial

Source: Screenshot generated in Discord

4
 Upon accessing, there will be a step-by-step description and the link to a
Google form, which must be answered, it doesn't ask for much.
information, just the email that you want to receive the priority, nickname of
Discord is a name
Once this is done, just wait for the email that will notify you that it is already
possible access.

4.2. Presentation of the Site

With access in hand, one must again access the main site and
log in with Google using the previously entered email.
Once logged in, the site will ask you for the username that
wants to use on the site, some areas of interest to improve a 'feed'
that the site has.
On the 'Home' tab you can already see the 'Featured Models', which are like
separate intelligences that generate different image styles even
using the same command prompts, these will be commented on and
detailed when the 'Finetuned Models' tab is explored.
Additionally, there are also the 'Featured Creations', where it is possible to
artworks generated that are trending at the moment ("Trending" tab),
it is also possible to see the latest artworks that have been generated (tab "New") and
also the most liked artworks on the site (tab "Top"), there is also a
search mechanism that can be used to search for specific arts that
may have already been created. In the 'Community Feed' tab, you basically have
even from the "Featured Creations" section of the "Home" tab.
Figure 4: Home and Community Feed Tab

Source: Screenshot generated in [Link]

5
 On the 'Personal Feed' tab, you can see your latest creations.
using AI, you can also see the artworks you liked as
also a feed with the creations of the profiles you followed on the site.
Figure 5: Personal Feed Tab

Source: Screenshot generated in [Link]

On the 'Training & Datasets' tab, it is possible to create image sets.

to train artificial intelligence and achieve eventual better results
in your creations, in which you can upload up to 40 images and
AI will process these images and learn common traits, style, etc.
AI training and the training time will vary depending on the
images that you upload, heavier and more realistic images will have
a higher computational cost.
Figure 6: Aba Training & Datasets

Source: Screenshot generated in [Link]

6
 Finally, we have the 'Finetuned Models' tab, where we can see
various options for art creation models, which function as
predefined datasets with a specific image style, where the
even prompts generated with different models will result in styles
different.
Figure 7: Aba Finetuned Models

Source: Screenshot generated in [Link]

Depending on what you want to achieve, some models are more

more favorable than others, such as, for example, the 'Cute Characters' which works
better for 3D characters in 'Disney' style, or the DreamShaper that
it can be more versatile in relation to environments and characters, but not
can be so ultra-realistic.
Figure 8: Some Finetuned models

Source: Screenshot generated in [Link]

7
 In addition to those provided by the site itself, there are also the models
made by the community, based on training by Datasets from others
users including you.
Figure 9: Community models

Source: Screenshot generated in [Link]

5. Creation of artworks using AI Image Generation

To start creating images, you should go to the tab
"AI Image Generation", which will open a menu with various options for
configure and thus shape your creation, which we will explore ahead:
Figure 10: Image creation area

Source: Screenshot generated in [Link]

8
[Link] System
The site has a Token mechanics, which works like a
currency that is spent when generating an image, daily you
starts with a total of 150 tokens, which can be freely spent until
At the end of the day, when reaching the time of the day's turn, all your tokens are
refunded back to the initial amount.
Figure 31: Remove Background

Source: Screenshot generated in [Link]

The cost per image is variable and depends on the configurations.

chosen, for example, both the 'Prompt Magic' function and the function
"Image Prompt" applies a multiplier of two times the value
consumed by image, the chosen resolution can also apply a
modifier on the value of token spent per image, so it should be taken
quite careful in the choice of configurations.
Figure 32: Token Cost Demonstration

Source: Screenshot generated in [Link]

9
5.2. Options sidebar:
[Link] of Images

On the left we have several options to create our images, such as

Number of Images, where we choose how many images we are going to generate.
start from the same prompt.
Figure 11: Number of images

Source: Screenshot generated in [Link]

5.2.2. Image Dimensions

We also have the 'Image Dimensions' section, which we can
choose the size of the images, scales, etc. Depending on the size
from the image, the number of images is limited due to the expense
computational, for example: from 726 x 726 the maximum number of images
generated is 4 (four).
Figure 12: Size of the images

Source: Screenshot generated in [Link]

10
 5.2.3. Prompt Magic & High Contrast
In addition, there is the "Prompt Magic" function, which makes the AI
better optimize the output of the inserted prompt, use this option multiplies
the token cost of its creation in 2 (two), the token system will be
explained further on, the impact of this option is so great that it is
recommended to keep the same always active, here is an example of
generated images where the option above is on and below is off
following the same command prompt:
Figure 13: Example of Connected Prompt Magic

Source: Screenshot generated in [Link]

In this option, you can still adjust the strength as well.

prompt to allow the AI to create more or stay focused on the prompt that
you typed. Along with the 'Prompt Magic' option, there is the option to
High Contrast, which creates images with more shadows and darker.
Figure 14: Prompt Magic and High Contrast Menu

Source: Screenshot generated in [Link]

11
 5.2.4. Guidance Scale and Tiling
Further down, we have the option of 'Guidance Scale', which controls the
how much the image generation process follows the inserted text
prompt, being controllable and varying the strength between 1 and 20, however, values
very high or very low can end up causing irregularities in
images even if the 'Prompt Magic' is enabled, then it is
recommended by the site to use with the value of 7 (seven).
Figure 15: Guidance Scale Function

Source: Screenshot generated in [Link]

The Tiling tool is useful when you need to create images that
There are patterns repeating, like backgrounds and things of that nature, for
to function more effectively, it is ideal to have a term in the prompt
pattern tiling
Figure 16: Tiling Function

Source: Screenshot generated in [Link]

5.2.5. Image Prompt, Image to Image & ControlNet

Below the 'Tiling' function, we have a panel with two options,
being these the 'Image Prompt' and the 'Image to Image', which allow us
direct the creation, but without the use of datasets.
Starting with the 'Image Prompt', which allows you to insert images
from your file explorer, these will guide the AI and show with the
creation should resemble the end of the process, this applies a multiplier

12
twice the number of tokens per image. It is still possible to change the
"Image Weight", which makes the creation look more like the image
as a reference, it is also possible to change the 'Prompt Magic Strength', which
modify how much the 'Prompt Magic' function also influences creation.
Figure 17: Image Prompt

Source: Screenshot generated in [Link]

On the other hand, we have the 'Image to Image' function that uses a
image as a base and changes based on the inserted prompt, where possible
change the 'Init Strength', which allows you to control how much will be preserved of the
main image.
Figure 18: Image to Image

Source: Screenshot generated in [Link]

13
 The 'Image to Image' function also enables 'ControlNet', which
influence the creations based on factors from the inserted image in the 'Image
to Image”, such as the pose of the image, the depth, and the contours,
it is also possible to use the 'ControlNet Weight' to regulate the weight of the
how much these factors will influence.
Figure 19: Control Net

Source: Screenshot generated in [Link]

5.2.6. Seeds
Finally, in the 'Advanced Settings' section, we have the 'Fixed' option.
"Seed" which we can use the seed used to generate another.
image as a basis to generate other images, very useful for generating
images from others, like in 'Image to Image'.
Figure 20: Fixed Seed Function

Source: Screenshot generated in [Link]

It can also be used to optimize creations iteratively.

as follows: generate a series of images with a prompt, choose the
but more promising is to copy the seed, add it to the 'Fixed Seed' and generate
again, finally repeat this process until you are satisfied
with the result.

5.3. Prompts, Negative Prompts and Prompt Generation

At the central top there is the command prompt input section,
what is the text that the artificial intelligence will base itself on and create the art,
conducting a vast search using neural networks. It is possible to insert the
prompt in any language, but it is more accurate for prompts in language

14
English. It is imperative to emphasize that the more precise the
description, the result of the image will be better, so do not hesitate to
insert details. There is still the option of 'Negative Prompt', which allows you
add a secondary prompt of elements that you do not want
your art, which the AI will contour.
Figure 21: Prompt Menu

Source: Screenshot generated in [Link]

Below the prompt space we have the Finetune Model that will be
used to generate the image, where there are some already pre-established and
also the option to choose a custom one or from the community.
Figure 22: Models Menu

Source: Screenshot generated in [Link]

15
 Finally, we have the option of 'Prompt Generation', where the AI will generate
detailed and complex prompt suggestions (ideal for images)
necessary) based on a basic prompt typed by the user, interesting
when you have ideas, but don't know how to describe them to the AI,
However, be careful, you only have the right to 1000 Prompt Generations.
Figure 23: Prompt Generation Menu

Source: Screenshot generated in [Link]

[Link] Buttons

When creating your image, hovering over some options

additional options will appear, including options to remove the background (transform
in PNG), perform Upscale, etc. These will be described in the next topics
with more details.
Figure 24: Additional buttons

Source: Screenshot generated in [Link]

16
 [Link] and Remove Background

The 'Unzoom' function is quite useful for expanding a creation

if you eventually feel that you are in a very closed plan, the
artificial intelligence will work to fill in the rest of the image without the
"zoom". For example, in the image below on the left we have the image
On the right, we have the "Unzoomed" image.
Figure 25: Comparison of Unzoom

Source: Screenshot generated in [Link]

There is also the 'Remove Background' function, which does not require
many explanations, just remove the background from the image using AI, only if
One must be careful with very realistic images with too many details.
"fluttering", as regardless of the tool used, it is very
It's complicated to remove the background from these images.

Figure 26: Remove Background

Source: Screenshot generated in [Link]

17
 5.3.3. Upscaling

Image upscaling is the process of increasing the resolution or

size of a digital image. It is used when an image of
low resolution or small size needs to be enlarged to a
higher version or higher quality.
The goal of upscaling is to improve the visual appearance of the image.
expanded, adding details and maintaining sharpness as much as
possible. This can be done through methods like interpolation, which
adds pixels between the existing ones, or super-resolution, which uses
algorithms to reconstruct the image at a higher resolution with
based on the original content.
In this tool, due to generating images of a maximum of 1024 x 1024
in size, provides us with the option to perform three types of Upscaling,
that have some differences among themselves that will be discussed
individually. It is also important to highlight that all options have
a cost of 5 tokens.
Figure 27: Example of Upscaling

Source: Image available at:

This new AI system could be science fiction but it
s-very-real-indeed

Starting with 'HD Smooth Upscaling', an upscaler that works.

well with unique and focused objects, but not suitable for images with
many fine details, as smoothing may end up correcting them.

18
 Figure 28: Smooth Upscaling

Source: Screenshot generated in [Link]

We also have the 'HD Crisp Upscaling', a great upscaler.

balanced that retains a good amount of detail and sharpness of the
original image.
Figure 29: Crisp Upscaling

Source: Screenshot generated in [Link]

19
 Finally, in the free plan we have the 'Upscale Image Alternate', which
in addition to improving the resolution, it redoes some details of the original image,
as an "alternative" image.
Figure 30: Alternative Upscaling

Source: Screenshot generated in [Link]

20