SmartSplat: Feature-Smart Gaussians for Scalable Compression of
Ultra-High-Resolution Images
Linfei Li1 · Lin Zhang1* · Zhong Wang2 · Ying Shen1
1School of Computer Science and Technology, Tongji University
2Department of Automation, Shanghai Jiao Tong University
Smarter Gaussians, Smaller Images!
Raw PNG: 10848 × 16320, 189 MB
Disclaimer
- Below, we demonstrate the advantages of SmartSplat in storing and transmitting ultra-high-resolution images, where it can compress images from hundreds of megabytes to just a few megabytes while maintaining high visual quality.
- Similar to JPEG, SmartSplat can provide varying compression ratios and qualities depending on the scenario.
- To achieve efficient rendering of Image Gaussians on WebGPU, the visualization NPZ files simplify the data by converting float32 attributes to float16 precision, thereby enhancing rendering speed and reducing file size; however, this results in discrepancies between the data format and the quantization strategy described in the original paper.
- You can click on the image cards to view the Image Gaussians. Due to network conditions (such as VPN), loading may be slow—please be patient!