GENERATIVE AI FOR CREATING HYPER-REALISTIC 3D HUMAN MODELS

Abstract

Generative artificial intelligence (AI) have by far been instrumental in building hyper realistic 3D human models, ensuring their usage in gaming, virtual reality and digital content creation. In this research, we explore techniques of predicting highly accurate and lifelike 3D human avatar using state of the art neural networks approaches. That is to say, the trend of generating high fidelity 3D representations using 2D image collections has been shown possible with recent works like AG3D and Get3DHuman. LumiGAN introduces relightable face generation to help realistic motion of a rendered person in dynamic lighting environment, and SMPLpix brings neural avatar to enable seamless human motion synthesis. This work integrates methods such as the Skinned Multi Person Linear (SMPL) model and 3D Morphable Models (3DMM) to improve the quality of pose, texture and geometry estimation to produce better quality facial and body reconstructions. The research also applies pixel aligned reconstruction priors, to improve spatial consistency and structural precision; and volumetric convolutional neural networks (CNNs) to enhance the quality for the spatial consistency and structural precision. Experimental results show better performance such as tending to generate photorealistic textures, photorealistic dynamical facial expressions, and human poses with little input data. The research promises to further enhance the experience in the area of augmented and virtual reality immersive experiences through improvements in these advances. This serves as future work that will deal with challenges to performance optimization in real time and cross-domain adaptability. This work is a contribution to trends in 3D model synthesis from concept to application, that fills the gap between computer vision, machine learning, and visual computing.