# 3D Scene Generation

Stable Virtual Camera is a 1.3B parameter general diffusion model developed by Stability AI, belonging to the Transformer image-to-video model. Its importance lies in providing technical support for novel view synthesis (NVS), capable of generating 3D-consistent new scene views based on input views and target cameras. The main advantages are the ability to freely specify the target camera trajectory, generate samples with large perspective changes that are smooth over time, maintain high consistency without additional neural radiance field (NeRF) distillation, and generate high-quality seamless loop videos up to half a minute long. The model is only freely available for research and non-commercial use, aiming to provide innovative image-to-video solutions for researchers and non-commercial creators.

HoloDreamer is a text-driven 3D scene generation framework capable of producing immersive and view-consistent fully enclosed 3D scenes. It consists of two fundamental modules: stylized rectangular panoramic generation and enhanced two-phase panoramic reconstruction. This framework first generates high-definition panoramic images as a complete initialization for the 3D scene, then quickly reconstructs the 3D scene using 3D Gaussian scattering (3D-GS) technology, resulting in view-consistent and fully enclosed 3D scene generation. HoloDreamer's main advantages include high visual consistency, harmony, and robust reconstruction quality and rendering.

WonderWorld is an innovative 3D scene expansion framework that allows users to explore and shape virtual environments based on a single input image and user-specified text. Through fast Gaussian voxel and guided diffusion depth estimation methods, it significantly reduces computing time and generates geometry-consistent expansions, resulting in 3D scene generation times of less than 10 seconds. It supports real-time user interaction and exploration. This opens up possibilities for rapidly generating and navigating immersive virtual worlds in fields like virtual reality, gaming, and creative design.

Text2Immersion is an elegant method for generating high-quality 3D immersive scenes from text prompts. Our proposed pipeline first uses pre-trained 2D diffusion and depth estimation models to gradually generate Gaussian clouds. Next, the Gaussian clouds are refined, interpolated, and enhanced to add detail to the generated scenes. Unlike mainstream methods that focus solely on individual objects or indoor scenes, or those that use downscaled trajectories, our approach can create a variety of scenes containing different objects, even extending to the creation of imaginative scenarios. Therefore, Text2Immersion can have a broad impact on various applications, such as virtual reality, game development, and automated content creation. Extensive evaluations demonstrate that our system outperforms other methods in rendering quality and diversity, and continues to advance text-based 3D scene generation.