DreamSurfels: Generative Gaussian Surfels for Geometrically Accurate 3D Content Creation

The field of 3D content creation has seen significant progress through optimization-based methods, such as score distillation sampling (SDS). Recently, DreamGaussian attempted to combine SDS sampling and 3D Gaussian splatting to accelerate the generation process but faced limitations in achieving precise geometric accuracy. This paper introduces DreamSurfels, an improved 3D content generation framework that enhances efficiency and geometric accuracy. We replace the 3D Gaussian Splatting in the original DreamGaussian with geometrically accurate Gaussian Surfels. This approach achieves precise radiance fields with cleaner normal and depth reconstructions. The proposed method maintains the rapid generation capabilities of the DreamGaussian, producing high-quality textured meshes from single-view images in just 1 minute, while delivering substantial improvements in geometric fidelity. Experiments validate the performance of DreamSurfels, making it a robust solution for efficient and accurate 3D content generation. These improvements pave the way for more practical applications in industries such as game design and digital media