Reproduce “LSNIF: Locally-Subdivided Neural Intersection Function” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shin Fujieda, Chih-Chen Kao, Takahiro Harada
LSNIF introduces a neural implicit representation that replaces traditional bottom-level Bounding Volume Hierarchies (BVHs) for ray-geometry intersection in real-time path tracing, achieving up to 106.2x memory reduction compared to compressed BVHs and supporting all ray types through offline per-object training.
Key insights
- Transforming irregular BVH traversal into regular neural network inference leverages GPU matrix cores more effectively.
- Providing detailed, localized ray-geometry interaction points (via voxelization) as input significantly improves neural network accuracy for intersection queries.
- Pre-training neural models on individual objects allows for compact, reusable representations that support dynamic scene elements and reduce overall memory footprint.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.