Reproduce “SmartSplat: Feature-Smart Gaussians for Scalable Compression of Ultra-High-Resolution Images” 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

Linfei Li, Lin Zhang, Zhong Wang, Ying Shen

Researchers from Tongji University and Shanghai Jiao Tong University developed SmartSplat, a feature-driven 2D Gaussian image compression framework that robustly handles ultra-high-resolution images. It achieves superior reconstruction quality at extreme compression ratios (e.g., up to 3000x on 16K images), demonstrating faster convergence and enhanced scalability compared to prior Gaussian Splatting methods.

Key insights

  • Intelligent, feature-guided initialization of Gaussian parameters (position, scale, color) is crucial for achieving high compression ratios while maintaining perceptual quality in UHR images.
  • Combining variational sampling for high-feature regions with exclusion-based uniform sampling ensures comprehensive coverage while efficiently allocating limited Gaussian resources.
  • Adaptive, scale-aware color sampling significantly improves reconstruction fidelity compared to simple random or pixel-center color assignments for Gaussians.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.