A Query-Efficient Stochastic Volume Rendering Framework for Time-Varying Implicit Neural Volumes

Time-varying implicit neural representations are becoming a practical way to store dynamic volumetric data, but interactive rendering has been a bottleneck because every sample along a ray can require a neural network inference. That makes the usual dense ray-marching playbook far too expensive for real-time use, especially when the volume changes over time.

The new framework tackles that problem with stochastic volume rendering built around delta tracking. Instead of brute-force sampling, it uses a four-stage pipeline that mixes heterogeneous parallelism: ray tracing cores handle traversal, while tensor cores batch the neural evaluations. The system also adds ray budgeting and query pruning to reduce the number of INR lookups per frame.

For developers working with neural rendering or scientific visualization, the practical takeaway is that many time-varying INRs can now be rendered directly from their original representation without resampling, caching, or retraining. On an RTX 4090, the renderer reaches roughly 30-40 FPS at 1024x1024 and converges to high-fidelity images, with timestep updates taking about 1-2 ms.

The broader significance is that the same techniques could inform real-time pipelines where neural inference has replaced cheap texture or volume fetches. The work sits at the intersection of graphics and machine learning, but the performance ideas are relevant anywhere query cost dominates frame time and temporal coherence can be exploited.