Nicolas Scheer
home / projects / bvhnet
2026Rendering enginev1 done

BVHNet

Physically based path tracer, on CPU (OpenMP) and GPU (OptiX / RT cores), importing glTF scenes exported from Blender.

Aperçu — BVHNet

Problem

Understand and control every stage of a path tracing pipeline — BVH construction, intersection, sampling — with the longer-term ambition of replacing the traversal heuristic with a small neural network (Neural BVH).

Approach

Binary BVH built by centroid sorting (simplified SAH), Möller–Trumbore intersection, path tracing integrator with Next Event Estimation and Russian roulette, PBR GGX/Cook-Torrance materials, ACES tone mapping. Full GPU port via OptiX to exploit hardware ray tracing.

Results

Working GPU rendering on complex glTF scenes (RTX 5070 Ti). The Neural BVH part remains a stated research goal, not yet implemented.

view code →

Other projects

all projects →