ConsistentPrecision-GS: Per-Gaussian Mixed Precision and Hardware-Backprop for 2–4× Faster Training

TL;DR: Use sharper math for hard parts and cheaper math for easy parts, decided by how well different views agree, to train way faster without getting blurry. We integrate a differentiable hardware rasterizer and choose per-Gaussian forward/backward precision (8/16/32-bit) based on multi-view consistency and gradient variance. First experiment: on RTX 4080s, measure full-pipeline speedups vs. tile-based software rasterization while matching FastGS quality.

Research Question: Can multi-view consistency guide mixed precision and hardware-accelerated gradient computation to speed up training substantially without degrading radiance field quality?

Hypothesis: Per-Gaussian/tile precision selection (unorm8/float16/float32) coupled with programmable blending and hybrid gradient reduction will yield 2–4× end-to-end training acceleration with negligible quality loss; the consistency signal will safely down-precision most Gaussians early and focus 32-bit only on unstable regions.

Experiment Plan: - Setup: Replace tile-based renderer with a differentiable hardware rasterizer; implement programmable blending with quad-level + subgroup reductions; design a scheduler that maps FastGS consistency + gradient variance to precision tiers for forward/backward per tile/Gaussian.

• Data/Materials: Mip-NeRF 360 and Deep Blending; RTX 4080 GPUs; identical training hyperparameters to FastGS where possible.
• Measurements: End-to-end wall time (forward+backward), GPU memory, PSNR/SSIM/LPIPS, precision distribution over training, ablations on precision policies and float16 vs. unorm16.
• Expected Outcomes: 3× wall-time acceleration (forward+backward) in line with hardware-rasterizer findings; quality parity with FastGS; evidence that consistency-aware precision avoids gradient degradation.

References: ['Ren, S., Wen, T., Fang, Y., & Lu, B. (2025). FastGS: Training 3D Gaussian Splatting in 100 Seconds.', 'Yuan, Y., & He, Q. (2025). Efficient Differentiable Hardware Rasterization for 3D Gaussian Splatting. arXiv.org.', 'Hanson, A., Tu, A., Lin, G., Singla, V., Zwicker, M., & Goldstein, T. (2024). Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives. Computer Vision and Pattern Recognition.', 'Kerbl, B., Kopanas, G., Leimkuehler, T., & Drettakis, G. (2023). 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics.']