Fast Neural Representations for Direct Volume Rendering

Despite the potential of neural scene representations to effectively compress 3D scalar fields at high reconstruction quality, computational complexity training and data step using representation networks limits their use in practical applications. In this paper, we analyse whether can be modified reduce these limitations such architectures also used for temporal tasks. We propose a novel design GPU tensor cores integrate seamlessly into on-chip raytracing kernels, compare quality performance network alternative network- non-network-based compression schemes. The results indicate competitive our rates, significantly faster decoding times lower memory consumption during reconstruction. investigate how density gradients computed show an extension where density, gradient curvature are predicted jointly. As spatial super-resolution approaches time-varying fields, solution that builds upon latent-space interpolation enable random access arbitrary granularity. summarize findings form assessment strengths domain volume rendering, outline future research directions. Source code: https://github.com/shamanDevel/fV-SRN