3D Avatar Diffusion neural network developed by Microsoft Research

The AI-powered tool creates a highly detailed 3D avatar viewed in 360 degrees based on just one photo and a short text description.

The generated avatar can be viewed in 360 degrees in high quality. The model performs impressive generation diversity in terms of gender, age, expression, face accessories. The final 3D model can be also edited by changing the skin color, hairstyle and adding accessories.

In order to optimize the memory and processing costs in 3D modeling Microsoft team proposed the roll-out diffusion network (Robin).

The generative model “represents a neural radiance field as multiple 2D feature maps and rolls out these maps into a single 2D feature plane within which we perform 3D-aware diffusion. The Rodin model brings the much-needed computational efficiency while preserving the integrity of diffusion in 3D by using 3D-aware convolution that attends to projected features in the 2D feature plane according to their original relationship in 3D.”

The 3D Avatar Diffusion neural network would be useful in creating a virtual reality (VR) or augmented reality (AR).  So far, there is a demo version only, but Microsoft Research team promised to make the AI-powered tool available soon.