PowerGS: Display-Rendering Power Co-Optimization for Foveated Radiance-Field Rendering in Power-Constrained XR Systems

We compare seven methods using standard metrics (PSNR, SSIM, LPIPS) across two datasets (MiP-NeRF360 and Synthetic NeRF): 3DGS, MiniSplatting-D, MiniSplatting, LightGS, FR-Display-H, FR-Render-H (RTGS), and PowerGS-H.

The first two are dense methods with high quality and power usage. The next two are pruned methods but use uniform rendering. The last three apply foveated rendering (FR): FR-Display-H adjusts color to reduce display power, FR-Render-H prunes to reduce rendering power, and PowerGS-H co-optimizes both for the lowest total power.

For FR methods, we report quality in the foveal region, as these standard metrics target foveal perception and HVS quality is already aligned across eccentricities.

We also provide scene-wise visualization results for 3DGS, MiniSplatting-D, MiniSplatting, LightGS, FR-Display-H, FR-Render-H (RTGS), and PowerGS-H. We include results from MiP-NeRF360, Synthetic NeRF, Tank and Temples, and Deep Blending — covering a total of 20 scenes. (The bicycle scene is shown in the main paper and thus omitted here.)

Each scene visualization includes both rendering and display power numbers (rendering, display), as well as zoom-ins of the foveal and peripheral regions.