← ML Research Wiki / 2309.16653

DREAMGAUSSIAN: GENERATIVE GAUSSIAN SPLAT-TING FOR EFFICIENT 3D CONTENT CREATION

Jiaxiang Tang School of IST National Key Laboratory of General AI Peking University, Jiawei Ren S-Lab Nanyang Technological University, Hang Zhou Baidu Inc, Ziwei Liu S-Lab Nanyang Technological University, Gang Zeng School of IST National Key Laboratory of General AI Peking University (2023)

Paper Information
arXiv ID
2309.16653
Venue
International Conference on Learning Representations
Domain
computer vision, machine learning, computer graphics
Reproducibility
4/10
Contents

Abstract

https://dreamgaussian.github.io/"A photo of an ice cream" "A photo of a hamburger" * This work was partly done when interning with Baidu Inc. and visiting NTU S-Lab.

Summary

The paper introduces DreamGaussian, a novel framework for efficient 3D content creation that leverages generative Gaussian splatting. It addresses the lengthy optimization times associated with current 3D generation methods and proposes a two-stage approach combining mesh extraction and texture refinement. DreamGaussian achieves high-quality textured mesh generation from a single view image in approximately 2 minutes, outperforming existing methods in both speed and quality. The authors emphasize the progressive densification of 3D Gaussians, which leads to faster convergence. Experiments demonstrate significant improvements in generation efficiency while maintaining competitive quality, paving the way for practical applications in 3D content creation across various industries.

Methods

This paper employs the following methods:

  • Score Distillation Sampling (SDS)
  • 3D Gaussian Splatting

Models Used

  • Neural Radiance Fields (NeRF)
  • Stable Diffusion
  • Zero-1-to-3

Datasets

The following datasets were used in this research:

  • None specified

Evaluation Metrics

  • CLIP-similarity

Results

  • Generates high-quality textured meshes in just 2 minutes from a single-view image
  • Achieves approximately 10 times acceleration compared to existing methods

Limitations

The authors identified the following limitations:

  • Not specified

Technical Requirements

  • Number of GPUs: 1
  • GPU Type: NVIDIA V100 (16GB)

Keywords

3D content creation Gaussian Splatting mesh extraction texture refinement neural radiance fields score distillation sampling diffusion models

Papers Using Similar Methods

External Resources

Funding: Not specified
References: 76
Influential Citations: 65