GAUDI: A Neural Architect for Immersive 3D Scene Generation

GAUDI allows modeling both dependent and independent distributions on complex 3D scenes. It can create 3D scenes with hundreds of thousands of representations for thousands of indoor scenes without mode collapse or canonical orientation problems during training. The new denoising optimization aims to find latent representations that jointly model the radiance field and camera pose separately. The approach obtains state-of-the-art generative performance on multiple datasets. The approach allows different generative settings, including unconditional generation and image- or text-based generation.