Neural Radiance Fields (NeRF) has shown great promise in the field of 3D scene reconstruction and view synthesis, offering an alternative to traditional geometric-based methods. Despite its success, NeRF suffers from several limitations, including high computational expense, slow processing times, and an inability to generate multiple 3D scenes. Recent work in novel 3D generation has shifted towards leveraging Generative Adversarial Networks (GANs) for synthesizing new 3D content. However, most existing approaches are either voxel-based or rely on NeRF for 3D scene representation, presenting a trade-off between performance and quality.

In this project, we explore the implementation of the EG3D - Efficient Geometry aware 3D GAN, a hybrid representation that combines a tri-plane encoding scheme for 3D features and geometry information with a compact Multi-Layer Perceptron (MLP) network for decoding. The tri-plane encoding allows for efficient storage and processing, while the MLP network enables high-quality neural rendering. The project focuses on the effectiveness of the tri-plane in encoding 3D features and its potential for improving neural rendering techniques. By analyzing the EG3D model, we aim to provide insights into the benefits and challenges of hybrid representations for 3D scene reconstruction and neural rendering, and contribute to the ongoing pursuit of efficient and high-quality 3D content generation.