“…However, the prospect of training a neural network, the added complexity of integrating machine learning inference with traditional graphics pipeline, and the proprietary nature of machine learning frameworks have stalled the industry-wide adoption of these techniques. The recent probe-based algorithm, Dynamic Diffuse Global Illumination (DDGI) [28], extending the classic irradiance probes, still remains an excellent choice due to its relative simplicity, quality, and cloud streaming capabilities [14,53]. However, scaling of DDGI in its original formulation is limited, and approaches such as multi-grid hierarchy and probe rolling [29] are necessary to scale it across large environments.…”