As part of our fruitful partnership with steepsoft to go the extra mile and infuse AI development to our clients' digital products where for better user experience and higher ROI, we suggested our friends at @MultiversX implementing AI-generated avatars. The challenge was in presenting the functionalities of this widely known phenomenon, yet merely understood, as well as developing it and implement it as part of the MultiversX/xPortal ecosystem.
We distilled its complexity to a straight forward proposition that translated in harnessing AI-generated images for the establishment of financial identities within the digital space. As financial transactions and services become increasingly digitized, the need for secure, user-centric solutions got imperative which is why we were sure that, for MultiversX, AI-generated avatars could play a pivotal role in this process, acting as a personalized interface for users to manage their finances and interact with various financial platforms.
Being presented with the challenge of personalizing experiences by implementing Avatars within xPortal, we anticipated that transformer-based systems were monumental in our research. A month’s worth of research has shown that using this architecture will lead to great results in creating hyper realistic avatars.
We figured that, once implemented, the real value of AI-generated avatars would transcend their current social media hype, redefining the way users interact with digital products and services, while opening up new frontiers for innovation and personalization in the digital world. We knew that we could use the momentum and provide a trigger for the early adopters of the DeFi space to enter this realm seemingly complex realm and learn to use it to their advantage, as well as coexist with the other users.
What we did
The path we chose to follow included a system based on Stable Diffusion, Dreambooth and other proprietary and non-proprietary pre/post-processing models. While Stable Diffusion is more than capable of generating hyper realistic images, fine-tuning the model on a set of selfies can easily become a challenge of great proportions, especially when integrated into a direct-to-consumer product.
We had to think of: low light detection, selfie angles and camera poses, background extraction, object detection, face detection, image enhancement, and “trainability” (how confident are we that it can be used for training purposes and not compromise results).
All these pre-processing tasks became necessary to provide a seamless and cohesive user experience, while post-processing tasks would guarantee avatar quality, avatar pose control and ensuring that unwanted objects are not generated.
In our case, choosing the right architecture, models and tools amounted to 30% of the complexity. xPortal has an impressive user base - hundreds of thousands of active users. Scaling hardware-hungry AI solutions to a community of this size is not a trivial task. We potentially needed 100 GPUs powered by high-end computing units, which can be scaled on demand, as users go through their avatar generation process.
As part of the process, we worked on answering integral questions such as:
- Which cloud computing provider could allocate GPU computing volume of this size?
- How do we scale hardware and on-demand availability within specific regions?
- How do we keep costs low, while also shortening wait time for users?
We chose to focus on using solely GCP for our training and inference pipelines, optimizing every step of the process and tailoring it specifically for high end NVIDIA GPUs.
Following a thorough Agile process based on continuous communication, we managed to successfully launch the platform on the 1st of March 2023, meeting a fairly tight time-frame.
We are going to work on improving processes to continue support the MultiversX team in reaching the milestone of onboarding 1 billion users to a new set of fun and useful services, experiences, and applications, making digital finance, Web3, and the Metaverse accessible and easy to use to anyone, anywhere.
Over the past decade, we have perfected our end-to-end development process to cover the entire product life cycle.