DGQC: Gradient-Driven Point Cloud Simplification for Efficient Gaussian Articulated Models

We present Dynamic Gradient Quantile Control (DGQC), a training-time optimisation framework that markedly reduces the computational footprint of Gaussian Articulated Template (GART) models. DGQC learns to adaptively sparsify the Gaussian point cloud by analysing gradient-magnitude statistics and applying a quantile-decay rule that preserves perceptually salient regions while pruning redundancy elsewhere. In contrast to fixed-threshold heuristics, our self-tuning strategy eliminates manual hyper-parameter search and scales gracefully with scene complexity. Evaluated on the PEOPLESNAPSHOT benchmark, DGQC cuts Gaussian count by 51.5% and raises render throughput by 4.6% relative to the GART baseline, with negligible quality loss (∆PSNR < 0.05 dB). These gains make articulated human capture feasible in real-time pipelines that demand accuracy and speed.