Features
- Efficient 3D Scene Representation: SceneDreamer employs a bird’s-eye-view representation with efficient height and semantic fields, offering both disentangled geometry and semantics for remarkable realism and complexity management.
- Generative Neural Hash Grid: Harnessing the power of a neural hash grid, SceneDreamer parameterizes the latent space, allowing it to encapsulate universally applicable features across scenes, effectively aligning content based on scene semantics and 3D positioning.
- Style-Modulated Renderer: With a style-modulated renderer, SceneDreamer masterfully blends latent features to produce photorealistic 2D images through the art of volume rendering, delivering visually stunning results.
Use Cases:
- Limitless 3D Scene Generation: Empower SceneDreamer with random noise as input, and witness the creation of expansive 3D environments with diverse landscapes and the freedom of unrestricted camera movement.
- 3D Scene Synthesis from Everyday 2D Images: Drawing wisdom from everyday in-the-wild 2D image collections, SceneDreamer possesses the unique ability to craft vibrant and multifarious unbounded 3D realms, all without the need for cumbersome 3D annotations.
SceneDreamer stands as a beacon of excellence in the realm of 3D scene generation, showcasing its supremacy in crafting realistic and diverse boundless 3D landscapes that surpass contemporary methods.
