DeepFakeAI

A Web-Based DeepFake Service for All

Project by Feeno0 & kaiser 938

May 7th, 2024

Abstract
DeepFakeAI aims to grant anyone access to the power of deepfake technology, utilizing the
blockchain to create a robust, cloud-based service that can quickly generate clips according to
it’s users desires. These clips are generated through text submitted by the user and a selected
”Character” on the site, which can be one of the pre-selected characters curated by the service or
a character generated through a user’s own video and audio data. Available through a web app,
web API or telegram bot, the service allows the generation of as many clips as users require.
Table of Contents
Table of Contents 1

Background 2

Enter DeepFakeAI: A One Stop Service for Deepfakes 2

Use-Cases for DeepFakeAI Clips 2

Promotion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
Entertainment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

The Generative Process 2

Character Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Curated Characters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
User-Generated Characters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Clip Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
HD Lip-syncing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

Accessing the Service 4

The Web App . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
The Web API . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
The Telegram Bot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

Payment Structure 5
Using Minutes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
User vs Developer Minutes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
Custom Characters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
Holding $FakeAI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
Tier Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

Road-map 7

Summary 7

1
Background
AI generated video clips created through deep learning, or deepfakes as they’re commonly known,
have become widely recognized as a formidable technology in our current world. Anyone can generate
clips so long as they have a powerful computer, several hours to train and refine a model, several
more hours to render generated clips and the money to foot the power bill. These requirements create
a high barrier to entry, and although some services exist to create or alter images, generate audio,
and even synthesize new clips, the process is often unpredictable, expensive, time intensive or a some
combination of these three issues.

If a person wanted to create a clip of a person saying their chosen phrase, they would need to use a
combination of the above services, then bring these individual elements together using video editing
software and a deft hand. In an age where deepfakes are so present in the public consciousness, it is
still considerably difficult to create a convincing video clip.

Enter DeepFakeAI: A One Stop Service for Deepfakes

DeepFakeAI is a service which implements deep learning strategies to create fake video clips of various
”Characters” as determined by text input given by the user. These ”Characters” can be chosen from
the ever-expanding list curated by the DeepFakeAI team, which are trained for long periods of time
and produce the highest quality clips available, or can be generated from user provided video and
audio data to emulate any person the user would like. This data is put through the same deep learning
algorithm as the curated ”Characters”, but due to varied user input a perfect outcome cannot always
be guaranteed. With proper input however, a high quality ”Character” can be generated and clips can
be produced quickly and easily.

Use-Cases for DeepFakeAI Clips

Promotion
With the ability to generate clips of any number of people promoting a product, it is easy to imagine
the plethora of possibilities when marketing that utilise the service. For instance, DeepFakeAI itself has
used the clip generation process to promote itself to great effect, increasing it’s popularity many times
over through social media engagement and video clips1 . It can also save resources during the production
of promotional materials, as actors do not need to be hired for the creation of advertisements.

Entertainment
DeepFakeAI’s most common use-case is the creation of clips purely for entertainment’s sake. These
clips can be generated as jokes, used to spread a message in a humorous way, or even as announcements
on platforms such as Discord servers or social media sites. So long as the use-case follows existing laws
and regulations on computer generated content, the clips are free to be used in any capacity the user
may desire.

The Generative Process

The generative process can be broken down into the Character Selection Process, Clip Gener-
ation, and the lip sync process, which can either use the standard lip sync process as a part of clip
generation, or an enhanced HD Lip Sync process, which can enhance the end product. Each step
has been carefully planned to create the highest quality clips possible in a reasonable amount of time,
and the DeepFakeAI team is constantly working on improving each and every aspect of the process.
1 DeepFakeAI does not condone the use of the product to imitate real people without clarification and clear disclosure

that clips are generated using a deepfake service.

2
 Figure 1: The Character creation and deepfaking process for curated characters

Character Selection
First, the user must select a character. These can be the pre-generated Characters that the DeepFakeAI
team has created, or can be a user-generated character that the user uploads video and audio data to
create.

Curated Characters
The curated Character list created by the DeepFakeAI team contains a wide variety of characters,
the ever-expanding list of which can be found proudly displayed on the DeepFakeAI website2 . These
Characters are generated by the team using the process shown in Figure 1, and the deep learning is
referred to therein as the ”Deepfake Audio Training Algorithm” and the ”Deepfake Video Training
Algorithm”. Both visual and auditory data is extracted from publicly available footage, but each are
carefully edited as two separate inputs to reduce unwanted bad data. The key to the quality of these
curated characters is in the Review, Refinement and Retraining process, wherein the DeepFakeAI team
tweak variables and monitor the deep learning progress throughout it’s development. By removing all
unwanted noise and other speakers, the end product is clean and clear for use in the clip generation
process.

User-Generated Characters
User-Generated Characters are created in a similar fashion, but go through a faster and more auto-
mated training process to ensure quick product turnaround for users. They are then used to create
clips for that user, and allow more individualized results for those who may want it. The quality is
still high, but depends much more on user input data and is limited by the quality of video and audio
2 https://fakeai.io

3
provided. The data can be given by uploading 1 minute or less video footage of the desired Character
and some audio , which is then processed in the same manner described above. There is then a training
period, and you can submit your email to be notified when the training is complete3 .

Clip Generation
Once a Character is selected, the clip generation process can begin. The user inputs a message to their
chosen interface, whether it be the web app, web API, or Telegram bot (See ”Using the Service” for
more details). The input message is then fed into the trained Character model , which processes the
message into a video clip using generative AI processing based on the Character model. The output is
a realistic video clip including the Character moving and speaking in front of a background, and audio
that is synchronized with the visuals which match up with the input message. This deepfaked clip is
the main output from the deep learning model.

HD Lip-syncing
While the initial output from the deep learning model is more than sufficient for many purposes, the
DeepFakeAI team has developed a more advanced HD Lipsyncing Algorithm that uses further pro-
cessing to greatly improve the quality of mouth movements, expressions, and synchronization with the
generated audio. By implementing this optional algorithm, the clips generated appear very realistic,
and can be viewed in higher quality settings without suffering reduced believably. For an example, see
DeepFakeAI’s HD Lipsync Model Demo4 . Users can choose whether to use this model or not in their
clip creation menu.
By utilizing the above processes, DeepFakeAI produces high-quality clips in an efficient manner re-
gardless of the message or Character chosen.

Accessing the Service

DeepFakeAI can be accessed in three main ways: the web app, the web API, and the Telegram Bot.
Each provides a different user experience tailored to it’s specific platform. All of the above methods are
mediated through the web portal at https://fakeai.io, which allows all of the tools for DeepFakeAI
to be conveniently stored in one place.

The Web App

The web app is the main method through which users can access DeepFakeAI for direct clip generation5 .
Through this web app, users can validate their email address with the site and create an account, which
will allow access to all aspects of the service. This includes purchasing minutes, linking a Telegram
Account to use the Telegram Bot, generating clips, accessing previous generation history, generating
API keys to use the web API, and accessing a history of previous transactions. Most users will want
to access the Web App tab, which allows direct clip generation via choosing a Character and entering
a prompt. These clips are generated and shown directly in the app, and this process can be accessed
either by using minutes purchased by the user or by using videos granted by the tier of the user
$FAKEAI6 .
The web app can also be used to create custom characters by accessing the Custom tab. Through
the Create function, a new character can be created for a fee. Once created, these characters are
available to use the same way as any other character. Simply upload video and audio of the character
in question, pay the fee, and allow the Character to process. Once done, it is available any time under
the List section of the Custom tab. To create your own character, go to custom characters section
in the App and enter a new name for your character.Then go to the List section and select your
character. Upload the video (mp4) and audio (wav/mp3). Once that is done, add a description for
your character and save it. Once start processing is selected, you can enter an email to be notified
3 For more on accessing the feature, see the Accessing the Service section.
4 https://www.youtube.com/watch?v=6QejGbEXYxk
5 The web app is available at https://app.fakeai.io
6 See Payment Structure for more details.

4
once the training process is complete, which may take a few hours. Once completed, your character is
ready for use in clip creation.

The Web API

The Web API is an additional method through which the service can be accessed, an API which can
be embedded in a user’s website to grant access to the service directly for visitors. The Web API
works very similarly to the web app functionality described above, and can be available for users to
access freely on the website, or in a monetized fashion7 .

The Telegram Bot

The Telegram Bot works slightly differently than the above methods, as it is managed through the
web portal, but accessed through the app Telegram. Users in a Telegram group which contains the
DeepFakeAI bot can use text commands such as /say or /ask to generate clips through the bot once
they have purchased minutes through the web portal and linked their Telegram account. The bot can
also be added to a server by purchasing Developer Minutes, and these developer minutes can either be
used by the purchaser, or enable users in that Telegram group to use the bot as long as minutes are
available.

Payment Structure
As a service with no outside advertisements available or sponsors to listen to, DeepFakeAI is able to
deliver clips without any outside interference or data sharing whatsoever. Because of this, the service
requires funding for further development, maintenance costs, and the base cost of hosting the service
into perpetuity. As such, the service is paid, and can be accessed through either holding the corollary
token $FakeAI, or by purchasing minutes directly on the web portal.

Using Minutes
Purchasing minutes is the most common way to use DeepFakeAI, which can be done through the web
portal. Users must be signed up and logged into the website before using the service.
If the use of fiat currency is desired, users can sign in using their email, the Google API, or the
X API. Once logged in, users must go to the ”Buy” tab of the dashboard, which can be found at
app.fakeai.io. Then, the user can select either User or Developer minutes, and choose how many
minutes they would like to purchase (More information on the difference between User and Developer
minutes can be found below). Once the desired type and amount of minutes have been selected, the
user select the ”Buy” option, which will redirect them to a Stripe checkout page. Here, payment
information can be filled out, and once the payment has processed the user will have the purchased
minutes available for use.
Alternatively, if the user wishes to use cryptocurrency, they can log in using their MetaMask wallet.
Once this is done, users can follow the same process described above, however instead of being redi-
rected to a Stripe checkout page, the currency will be pulled directly from the selected wallet upon
purchase.

User vs Developer Minutes

Users can purchase User Minutes or Developer Minutes, with a few key differences between the
two options:
7 All clips must still be paid for either by using the videos granted by the tier of the user or by using Developer

Minutes, for more information see the section on Payment Structure.

5
 User Minutes Developer Minutes
$2 per minute $3 per minute
Web App Access ✓ Web App Access ✓
Telegram Bot Access ✓ Telegram Bot Access ✓
Add Telegram Bot to Server ✗ Add Telegram Bot to Server ✓
Web API Access ✗ Web API Access ✓

Custom Characters
Custom characters can be created for a fee, and can be paid for using either cryptocurrency or fiat
currency using the same process as the minutes purchasing above. Simply select the dollar sign by the
Custom Character’s name in the Custom tab to access the checkout page. Once paid for, the character
can be trained. Once training is complete you will be notified via email. Once trained, the character
can be used just like any other, and can be viewed and managed under the List portion of the Custom
tab.

Holding $FakeAI
Cryptocurrency users can also access the service by holding a certain number of $FakeAI tokens. This
gains access to certain tiers of use, all of which give the user access to the same options as the Developer
Minutes. The uses include the following for each tier:

Tier Structure
• Tier 1: 50,000 - 99,999 tokens
– 1 video per day

• Tier 2: 100,000 - 199,999 tokens

– 2 videos per day
• Tier 3: 200,000 - 299,999 tokens
– 3 videos per day

• Tier 4: 300,000 - 499,999 tokens

– 4 videos per day
• Tier 5: 500,000 - 999,999 tokens

– 8 videos per day

• Tier 6: 1,000,000+ tokens
– Unlimited videos

All tiered options require locking the tokens for at least 24 hours to be used, after which the benefits
are available so long as the tokens continue to be held.

6
Road-map
The current road-map for DeepFakeAI is the following as of the writing of this whitepaper:

• Improved deepfake quality through continued training and refinement of the deep learning algo-
rithms and processes
• Further optimizations to ease of use and efficiency of the web app, web API and Telegram bot
• Introduction of the DeepFakeAI bot to new platforms, to expand use-cases across multiple social
media sites
• Continued expansion of the Character pool with input from the DeepFakeAI community
• Improvement of the custom Character option through further refinement of the automated Char-
acter creation process and custom refinement for video and audio

• Access to HD options for Custom Characters

• An ”Explore” tab, which will allow users to explore other user’s Custom Characters
• Free Character creation for holders of $FAKEAI

• A new ”Zero Shot” model, which will reduce cost, amount of audio data required and training
time significantly for custom characters alike using a new audio training process. These savings
will be passed on to the users as lower cost and time to use the service.
These objectives continue to drive the DeepFakeAI team to improve the service, and there is no limit
to the possibilities in the future.

Summary
DeepFakeAI as a service has revolutionized the easy creation of deepfaked clips, allowing users to
generate any desired clip with minimal effort at low cost. The technology is sound, the team is
working hard on improvements, and the output is clean, high definition video available for whatever
use-case can be imagined. It lowers the high barrier of entry to this powerful technology, and the
customization options are unrivaled by other competing services. All that is required is a computer,
an internet connection, and an idea to start creating.