Is it possible to use artificial intelligence to create hyper-realistic avatars based on uploaded selfies?

CodeLink identified a gap in the market for creating hyper-realistic avatars, given the current quality and cost of competitor products. To bring a new level of quality and realism to the avatar creation market, CodeLink wanted to leverage the skills of its internal R&D team. Additionally, they aimed to improve processes to reduce computation and user costs.

The CodeLink internal team worked autonomously to design and build the Android and iOS mobile applications.

The project kicked off in March of 2023 and is still ongoing.

To develop the app, we started by thoroughly evaluating the quality of AI-generated avatars using Stable Diffusion and Dreambooth. We used different prompt engineering practices to add styles like Samurai, Victorian, Viking, Vampire, and more. We optimized the model for training, inference, and avatar quality, and utilized various Google Cloud Platform services to host the training and inference servers. Our team also experimented with Low-Range Adaptation (LoRA) to reduce training time and decrease the size of training artifacts from 7GB to only 12MB. The final app generates realistic avatars that retain users' facial features while transforming them into different styles.