My3DGen is a system designed to create a personalized and lightweight 3D generative model using as few as 10 images. The system can reconstruct multi-view consistent images from an input test image and generate novel appearances by interpolating between any two images of the same individual. The approach involves utilizing a pre-trained model with fixed weights as a generic prior, while training a separate personalized prior through low-rank decomposition of the weights in each convolution and fully connected layer. To avoid overfitting, a regularization technique based on symmetry of human faces is introduced. The final system introduces only approximately 600,000 additional parameters per identity, achieving a 50-fold reduction in model size without sacrificing the quality of the generated 3D faces.

 

Publication date: July 11, 2023
Project Page: https://luchaoqi.github.io/my3dgen
Paper: https://arxiv.org/abs/2307.05468