DreamGaussian

DreamGaussian is a research framework presented at ICLR 2024 (Oral) by researchers from Peking University, NTU, and Baidu. It introduces a generative Gaussian splatting approach for efficient 3D content creation from a single-view image or text prompt. The method converges in 2 minutes for image-to-3D, approximately 10 times faster than existing optimization-based methods. It includes stages for generative Gaussian splatting and mesh texture refinement, outputting textured meshes. The system supports images with non-zero elevation angles and text-to-3D via a text-to-image-to-3D pipeline. Demo videos were recorded on an NVIDIA 3070 (8GB).