Offline GPU Path Tracer

Progressive spectral Monte Carlo GPU path tracer in Slang and C++ using Vulkan

This project is a GPU-accelerated offline path tracer combining modern acceleration structures, efficient instancing, and progressive rendering techniques. The renderer supports a mix of primitive types and large-scale repeated geometry while targeting high performance on the GPU.

GPU path tracer scene with transformed spheres — Transformed spheres and Gaussian splat content in the renderer.

Gaussian splat interior rendering — Gaussian splat scene rendering using stochastic ray tracing.

The full project write-up, source code, and latest updates are available on GitHub.

Project features

Sphere primitives
Triangle primitives with OBJ loading
Gaussian splat primitives with PLY loading
BSDF material support including IOR and metallic parameters
Mesh lights
CPU BVH construction and GPU traversal
Progressive path tracing with temporal accumulation
Transforms (location, rotation, and scale)
Instancing support
TLAS/BLAS separation
Mixed node types in the same tree, including transform nodes, primitives, binary AABB nodes, and 8-wide KDOP nodes
Nested instances and nested transforms
Arbitrarily large repeated-geometry scenes with reduced memory and build-time cost
Spectral rendering

What these features enable

The renderer is designed to handle complex scenes with repeated objects and large-scale instance hierarchies while keeping memory usage manageable. Transform and nested-instance structures allow a scene to be represented efficiently, making it possible to build scenes with many repeated groups of geometry without paying a full cost for each copy. The project also combines path tracing with temporal accumulation to refine the image progressively over time, creating high-quality images while still being responsive to real-time camera repositioning.

Implementation details

The project uses a Vulkan-based GPU pipeline with Slang for shader logic and C++ for engine structure and scene setup. Scene acceleration is organized around a hierarchy of instances and collections, and mixed node types help balance traversal efficiency and flexibility. Gaussian splat support is handled with stochastic ray tracing techniques to model transparency and alpha blending in a Monte Carlo framework.

References and related work

Sun, Xin, et al. “Stochastic Ray Tracing of Transparent 3D Gaussians”
Benjamin Feldman, “3D Gaussian Splatting in a Weekend”
Arman Uguray, “Ray Tracing: GPU Edition”
Vaidyanathan, Karthik, Sven Woop, and Carsten Benthin, “Wide BVH traversal with a short stack”

If you were looking for my CPU path tracer instead, you can find it here.