Reproduce “Atlas Gaussians Diffusion for 3D Generation” 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
Haitao Yang, Yuan Dong, Hanwen Jiang, Dejia Xu, Georgios Pavlakos, Qixing Huang
Researchers from The University of Texas at Austin and Alibaba Group developed Atlas Gaussians Diffusion, a framework integrating 3D Gaussian Splatting into the VAE + LDM paradigm for high-fidelity, efficient 3D content generation. This approach generates detailed 3D shapes with faster inference times, outperforming existing baselines in both unconditional and text-conditioned tasks.
Key insights
- The Atlas Gaussians representation efficiently scales the number of generated 3D Gaussians from low-dimensional latents without increasing network parameters, which is critical for generating fine details.
- A transformer-based decoder with disentangled geometry and appearance features, utilizing local self-attention with broadcasted global features, enhances learning effectiveness and computational efficiency.
- Using a learned nonlinear weight function for feature decoding provides superior representation capabilities compared to standard bilinear interpolation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.