The paper introduces DreamGaussian, a new framework for 3D content creation. It addresses the limitations of optimization-based 3D generation methods, such as slow per-sample optimization. The key feature of DreamGaussian is a generative 3D Gaussian Splatting model, which allows for efficient mesh extraction and texture refinement. The framework significantly speeds up 3D generative tasks and improves the quality of textures. It also includes an algorithm that quickly converts 3D Gaussians into textured meshes and refines details. Compared to existing methods, DreamGaussian is approximately ten times faster, producing high-quality textured meshes in just 2 minutes from a single-view image.
Publication date: 28 Sep 2023
Project Page: https://dreamgaussian.github.io/
Paper: https://arxiv.org/pdf/2309.16653
