Reproduce “GaussianSR: 3D Gaussian Super-Resolution with 2D Diffusion Priors” 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
Xiqian Yu, Hanxin Zhu, Tianyu He, Zhibo Chen
GaussianSR, developed by researchers at the University of Science and Technology of China and Microsoft Research Asia, enables high-quality, high-resolution novel view synthesis from low-resolution inputs by distilling 2D super-resolution diffusion priors into a 3D Gaussian Splatting representation. The method achieves superior visual quality and maintains real-time rendering speeds compared to prior Neural Radiance Field-based approaches, addressing challenges in high-frequency detail inference and rendering efficiency.
Key insights
- Large-scale 2D super-resolution diffusion models contain rich high-frequency details that can be effectively distilled into 3D representations for high-resolution novel view synthesis.
- 3D Gaussian Splatting provides a highly efficient alternative to Neural Radiance Fields (NeRF) for high-resolution novel view synthesis (HRNVS) while maintaining high quality.
- Direct application of generative priors like Score Distillation Sampling (SDS) to explicit 3D representations with adaptive optimization (like 3DGS densification) requires careful stabilization to prevent undesirable over-reconstruction and redundant primitives.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.