AMGSRN++: Improved Adaptive SRN for Scientific Visualization

We present AMGSRN++, which advances previous state of the art APMGSRN along three key directions. First, we implement efficient CUDA kernels to fuse the encoding operation into a single kernel, reducing VRAM requirement by over $50 \%$ improving throughput, enabling faster training and rendering with the availability for lower-end hardware to perform neural volume rendering efficiently. Second, we introduce a compression-aware training strategy for efficient feature grid compression when saving, reducing storage costs by $80 \%$. Lastly, we extend the method to time-varying data with a 3D+time approach, allowing parallel training with no dependence between timesteps for highly efficient model fitting. We extend the previously released rendering tool to support the new model, including seamless time-varying dataset visualization. As a result, time-varying datasets over 100 GB can be rendered in real time on consumer hardware with as little as 1 GB of VRAM, and using only 88 MB of storage space. Comparisons with state of the art compressors and other SRNs are provided, displaying continued strong representation capability and higher compressive capabilities. All code is released publicly at https://github.com/skywolf829/AMGSRN.