Grendel-FastGS: Distributed, Batched Multi-View Consistency for Scalable, Budgetless Training

TL;DR: Look at many camera views at once so you can tell what’s really solid, then cut the fluff earlier and train faster across many GPUs. We merge FastGS with Grendel’s multi-GPU, multi-view batching so consistency is computed over large batches, enabling stronger early pruning and stable densification. First experiment: replicate Grendel’s large-scene setups and test whether batched consistency improves both speed and final PSNR at scale.

Research Question: Does aggregating consistency over large multi-view batches in a distributed setup (sparse all-to-all) amplify FastGS’s gains and enable earlier, more confident pruning without quality loss on high-res, large-scale scenes?

Hypothesis: With more simultaneous views per iteration, consistency estimates become less noisy, allowing aggressive, budgetless pruning and targeted densification; coupled with sqrt(batch) rule scaling, this yields better PSNR-time trade-offs than single-GPU FastGS and improves scalability.

Experiment Plan: - Setup: Integrate FastGS’s importance computation into Grendel; compute per-Gaussian consistency over batched views and propagate these to partitioned pixel workers via sparse all-to-all; apply sqrt(batch size) scaling for learning rates and consistency thresholds; dynamic load-balancing keeps hot regions well-resourced.

• Data/Materials: Rubble dataset and other large-scale, high-resolution scenes; 8–16 GPUs; baselines: single-GPU FastGS and Grendel without consistency-aware pruning.
• Measurements: PSNR/SSIM/LPIPS; training wall time; scaling efficiency (strong/weak scaling); number of Gaussians vs. iteration; communication overhead; ablations on threshold scaling.
• Expected Outcomes: Higher PSNR at the same training time or equal PSNR faster; stable Gaussian counts without explicit budgets; strong scaling close to Grendel while improving quality/time trade-offs.

References: ['Ren, S., Wen, T., Fang, Y., & Lu, B. (2025). FastGS: Training 3D Gaussian Splatting in 100 Seconds.', 'Zhao, H., Weng, H., Lu, D., Li, A., Li, J., Panda, A., & Xie, S. (2024). On Scaling Up 3D Gaussian Splatting Training. International Conference on Learning Representations.']