Stability and Parameter Tuning for Incoherent-Ray Training

Investigate and determine stable training procedures and hyperparameter settings for optimizing particle-based radiance fields using incoherent (randomly sampled) rays, including resolving training instabilities observed on the Deep Blending Playroom scene.

Background

During evaluation on the Deep Blending dataset, the authors report a larger quality drop for the faster variant of their method when trained with incoherent rays, attributing this to observed training instability on the Playroom scene.

They explicitly defer a deeper investigation and parameter tuning of incoherent-ray training to future work, indicating that robust and generalizable training protocols under incoherent sampling remain unresolved.