FAQ

 

AI Avatar’s Magic Avatars Explained

<aside> 👉 This is an official page created by Awesome AI Labs, Inc. the company behind AIAvatar app. This is live FAQ, new sections might be added.

UPDATE

AIAvatar introduced NSFW content moderation solutions and updated Privacy Policy to make it easier to read and avoid misinterpretations. Users can now permanently delete avatars from their devices and our servers.

</aside>

How AIAvatar works?

• The Magic Avatars feature operates based on an open-source neural network model called Stable Diffusion. It is made by Stability AI.
• The Neural Network was trained on a sizable set of Internet data to generate images from the text prompts – small pieces of text describing the desired scene in the output image.
• By default, Stable Diffusion generates a random person's appearance if the prompt mentions a term “human” in the image.
• AIAvatar leverages a copy of the model to make output images personalized, and re-trains it on 10 to 20 user photos.
• Each copy is trained individually, meaning it learns specific features of a user’s appearance, it interacts with at a given time. For that precise reason our Terms of Use request the permission to "to use, reproduce, modify, distribute, create derivative works of your User Content." Without this permission, personalized Avatars are impossible.
• The training process takes on average 10 minutes and requires an enormous amount of computational power – the machine is making approx. 120 million billion (yes, it's not a typo) of mathematical operations to analyze a single set of photos.
• The trained model starts generating personalized images – Avatars – consuming approximately the same amount of computational power.
• After the batch of Avatars is created, the shared photos are deleted permanently from our servers, as well as the copy of the model. Each time a user purchases a new pack of Avatars, the process repeats from scratch; that's why we ask you to upload photos every time you request a new package.

Can the model be used for facial recognition systems?

• No, AIAvatar doesn't have a "one-size-fits-all collective neural network" trained to reproduce any face based on aggregated learnings.
• Each time the user uploads a picture, an individual copy of the model gets trained to reproduce one particular face. This means that every user will encounter an individual version of the model.
• Once the model is done with production of Avatars, it gets deleted along with the original photos.
• Each model takes 4Gb of disk space. Storing every copy for each of millions of our users is neither necessary nor sustainable for our business.

What data do we store and where?

• We are fully GDPR and CCAP compliant. We store the minimal amount of data necessary to provide quality service to our users. Our servers are located in northern Virginia, Ohio, and Oregon.
• The photos uploaded to create avatars are deleted from our servers after the processing is complete, along with the model trained on each individual set of photos.
• Users can permanently delete their avatars from the device and our servers by swiping left or the chosen pack at any given moment.
• For other AIAvatar services, we request the rights to store and process user content to develop the apps and add new features.
• By default, we don't have users' contact data or banking details. App Store and Google Play aren't sharing information about your account or purchases. The app and stores exchange so-called "receipt" or Transaction ID – a string of digits that serves as the key to confirm the success of the transaction but bears no information about the user.
• Unless a user creates a AIAvatar account within the app, we don't have access to or store any of their contact details. Each user for us is represented by Session ID, the code of the session on a particular device.

Do you sell my data?

• No. AIAvatar does not sell the data to third parties. We request your permission to manage the User's Content in order to upload the photos to the cloud providers (Google Cloud Platform and Amazon Web Services) and apply styles, filters, and effects or produce avatars for you. In any case, the photos become available to us anonymously and are automatically deleted within 24 hours after being processed by AIAvatar.
• We have updated our Privacy Policy to make it more comprehensive and exclude the possibilities of broad interpretation. Further details can be viewed in clause 5 of the Terms of Use and our Privacy Policy.

Who owns the avatars?

• The Magic Avatars created via AIAvatar belong to you, the user who uploaded their own selfies on the app.
• By running the AI on your pictures, you engage in a creative act comprising images that are considered your original creation. AI is just a tool in this instance.

How can I use AIAvatar in my artwork?

• Feel free to use images from AIAvatar Magic Avatars for anything, including commercial purposes.
• Please also note that the copyright of AI-generated images is a complex area to navigate and the rules will vary from jurisdiction to jurisdiction, but we believe that you are the true rightsholder in respect of your Magic Avatars, and you are the one to decide how to apply them.
• We encourage users to abstain from usage of Magic Avatars’ creations to enable any deceptive activities and malicious intents of any nature.

What's the relation between Avatars and Artists' works?

• Neural Networks learn to recognize specific patterns and connections between the images and their text descriptions, not individual artworks. This way, the AI develops a mental model, the general operational "how-to" principles, that can be broadly applied in the process of content generation.
• Once the training is finished, AI doesn't refer to the original dataset. Instead, it applies the acquired principles it has developed to the process of further creation. Hence the outputs are not replicas of any particular artist's artwork.
• The Stable Diffusion Model was trained on LAION 5B, an enormous uncurated dataset designed to serve as a general representation of the language-image connection on the Internet. There's an understanding within the industry about the necessity of offering artists the opt-out option, introduced at the level of the entity that performs initial training of the model. The license of the current model is authorized for commercial use.

Why do some pictures look like there is an artist's signature on them?

• The details perceived as signatures are observed in styles that imitate paintings. This subset of images, more often than not, comes with sign-offs by the author of the artwork.
• As a neural network analyzes the images corresponding to text prompts about paintings, it inevitably memorizes the detail that is omnipresent in a subset.
• The trained network recreates this detail to make the output look more like paintings. The artifacts you see in the outputs do not utilize any existing language; often, they don't feature any letters. And it is not the remains of existing artists' signatures. A similar effect is observed across Sci-Fi and Astronaut style imagery. It essentially mimics movie posters -the letters that make no sense to humans to AI look like the examples it was trained on.

Why do female users tend to get results featuring an over sexualised look?

• Please note, occasional sexualization is observed across all gender categories, although in different ways.
• The stable Diffusion model was trained on unfiltered Internet content. So it reflects the biases humans incorporate into the images they produce. Creators acknowledge the possibility of societal biases. So do we.
• Creators of the LAION 5B model introduced NSFW filters to reduce the bias. Stability AI has also introduced changes to the model, making it harder for users to generate NSFW content.
• AIAvatar introduced the adaptations aimed at reducing the risks of either encountering or generating NSFW content. We have tweaked parameters in the model to make it generate fewer inappropriate images and introduced the NSFW filter that would prevent such content from being uploaded on the app to prompt the AI's response.
• Should inappropriate content get generated despite the precautions, the third level of protection, NSFW-filter, would blur the avatars and show them only upon the users' request.
• Despite the update, AIAvatar’s Magic Avatars feature remains available to adult users only.
• The intentional use of the app to create explicit content is absolutely prohibited by our Terms of Use and obliges AIAvatar users to comply with that requirement.

How can I withdraw my picture from the LAION 5B?

• LAION 5b model data was intended to be a general representation of the language-image connection of the Internet so there is no opt-in or opt-out solution in place.
• For citizens of the EU there's a procedure to remove their personal data, if encountered, from the set. Under GDPR terms, the name corresponding with other identifiable data is considered to be personal information.
• In the future, for other models, Stability AI is creating an opt-in and opt-out system for artists that can be used by services deploying the model.

Why is my device not supported?

• The technologies we use in our apps can be quite demanding for devices' architecture.
• For Android, the app requires at least version 8 and processor arm64-v8a.
• For iOS, system requirements are 15 iOS and higher.

Where is Awesome Labs located?

The company is incorporated in the US and based in Cyprus, European Union.

Why the feature is paid if it is based on AI in the public domain?

• Training of the AI model and avatar generation requires tremendous computational power.
• Covering the computing costs is a challenge for all providers. Therefore, most of them offer limited free trials and paid tiers. Stable Diffusion might be free if you deploy it on a gaming PC with a powerful GPU and apply coding skills to set the system up.
• Besides computing, generative neural networks require writing a concise prompt to kick off the process. At AIAvatar, we have developed and tested hundreds of style prompts to achieve predictable results that work for most users.
• With the Magic Avatars feature, we charge users for easy access to the powerful contemporary AI, enabling anyone to interact with the neural network without needing software installation, heavy-duty local GPU, coding knowledge, and writing skills.