Reproduce “GVGEN: Text-to-3D Generation with Volumetric Representation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xianglong He, Junyi Chen, Sida Peng, Di Huang, Yangguang Li, Xiaoshui Huang, Chun Yuan, Wanli Ouyang and 1 more
Researchers from Shanghai AI Laboratory and collaborators developed GVGEN, a text-to-3D generation framework that directly produces structured 3D Gaussian representations in approximately 7 seconds per object. This method achieves a CLIP score of 28.53 for semantic alignment and generates high-quality 3D models with diverse outcomes by employing a novel GaussianVolume fitting and a coarse-to-fine generation pipeline.
Key insights
- Structuring 3D Gaussians into a fixed GaussianVolume, with centers optimized as offsets from grid points, makes this explicit representation compatible with generative neural networks while retaining detail.
- The Candidate Pool Strategy successfully adapts dynamic 3DGS-like densification and pruning behaviors to a fixed-count, structured Gaussian representation, enabling robust fitting.
- A coarse-to-fine generation strategy, starting with a coarse geometric representation (GDF), effectively manages the high-dimensionality of 3D Gaussian attributes, improving convergence and quality for the generative model.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.