DreamGaussian: Generative Gaussian Splatting for Efficient 3D Content Creation

Targeting image- and text-conditioned 3D asset creation, the paper argues that NeRF-based SDS optimization is slow due to volume rendering cost. It instead optimizes 3D Gaussians with SDS, using progressive densification suited to generative optimization, then extracts a mesh and refines textures in UV space. Experiments highlight large wall-clock speedups over prior lifting methods while reporting competitive quality for common text- and image-to-3D setups.