Efficient Volume Rendering via Precomputed Density Queries and Predictive Break Conditions for Neural Smoke Reconstruction

초록

This study proposes a learning-free optimization framework that combines a precomputed density query and a predictive break condition to reduce the bottleneck of exhaustive voxel traversal in large-scale volume rendering. In the initial frame, cumulative opacity and density prefixes are mapped to spatial coordinates and cached, while inter-frame gradient vectors are used to predict the termination point of the next frame in advance, enabling early decisions before ray accumulation (predict-then-skip). Implemented in PyCUDA, the proposed method achieves up to 20x speedup compared to conventional early termination, while maintaining a maximum density error of <= 0.03 relative to the original, thus preserving boundary sharpness and temporal continuity. Furthermore, the predictive signals (gradient/occupancy/termination threshold) are integrated into the NeuSmoke framework: in Stage 1, the Neural Transportation Field reduces redundant accumulation through the predict-then-skip scheme and dynamically adjusts the near-far range and sampling rate, whereas in Stage 2, the CNN-based detail refinement uses predictive feature maps (edges, residuals, and confidence) as auxiliary inputs to maintain visual fidelity while shortening inference time. As a result, both quantitative and qualitative experiments confirm that the proposed approach significantly reduces per-frame computational load and inference time while preserving iso-quality. The framework is broadly applicable to volumetric scalar fields such as smoke/fluid simulations, CT/MRI data, and meteorological 3D fields. Finally, standalone PyCUDA tests demonstrate up to 20x acceleration over traditional accumulation, and NeuSmoke integration achieves an average of voxel traversal -42% and inference time 0.75x compared to the original NeuSmoke.

키워드