CSE 168 Final Project

Introduction

Through this project, techniques from the NVIDIA ReSTIR Global Illumination paper will be implemented on top of the basic path tracer developed in CSE 168 with Monte Carlo path tracing, next event estimation, Russian Roulette technique, and importance sampling. While a basic path tracer samples paths independently per pixel, ReSTIR introduces path reuse across frames and across neighboring pixels. This reduces calculation costs per pixel and the combining of multiple, even low quality samples, can produce a high quality sample per pixel. This allows one to gain high quality renders even in resource constricted environments. Implementing this technique involves several steps including sampling multiple paths per pixel with unbiased contribution weights per candidate, implementing reservoir sampling to select one candidate, combine this selected reservoir with the previous frame’s reservoir, pulling in neighboring sample contributions, then accumulating all this information into a final pixel color. This results in each pixel having an output radiance, direction, contribution weight, and reservoir for use with neighboring samples and future frames. With this technique it is hoped that renders will be more computationally efficient and have less noise especially in resource constrained environments. The final project includes many features and optimizations on top of the path tracer built throughout CSE 168.

Final Hero Render with ReSTIR and Dielectric

Deliverables:

Refactoring Code
OptiX Denoiser
Environmental Maps
Object Importing
Texture/Normal Maps
Resampled Importance Sampling (RIS)
Dielectric Material
funLight Generation
ReSTIR

Refactoring Code

This phase involved shifting sampling methods, probability density functions, and helper code to separate functions for cleaner, initial Path Tracer code.

Denoiser

Renderer.cpp

This deliverable involved including/turning on the built-in deep learning denoiser provided with OptiX.

Environmental Maps

SceneLoader.cpp

This part of the code added in an environmental map which when a ray misses an intersection and defaults to background color, instead add in an environmental map to be used as background. This environmental map (sampled in hit program) provides realistic looking background graphics. Incorporating this required loading of maps in host code and passing to OptiX.

Object Importing & Texture/Normal Mapping

SceneLoader.cpp

This part of the project built compatibility with ".obj" 3D models in Path Tracer scenes.

Resampled Importance Sampling (RIS)

PathTracer.cu

Two versions of this sampling method were implemented with version 2 being used for ReSTIR. In version 1, this sampling method takes a generated sample based on a given BRDF (GGX, dielectric, Phong, etc.) and stores the generated direction as a candidate. The weight of this direction candidate is then calculated using the BRDF and light probability density function. These weights are stored in an array as well. The same process of generating sample directions and weight candidates is also then done across quadlights. After RIS sample candidates and their associated weights are generated from both the BRDF and quadlights, a candidate is randomly chosen for use, using the weights as probability distribution for selection. The contribution of the chosen candidate is scaled using a factor incorporating the number of samples as well as the BRDF's pdf. With version 2, rather than using generated direction for samples, points on light sources were used instead. This modification is required for compatibility for spatial reuse in ReSTIR as when using direction for sampling, one is not guaranteed to hit a light.

As can be seen in these images, the RIS render reduces overall variance compared to the other techniques. Yet, the MIS render is able to handle specular reflections better because our RIS is only using light sampling. The RIS render is very similar to the NEE render but using a lower sample count for the same quality

Dielectric Material

PathTracer.cu

This function models how light interacts with dielectric material supporting internal reflection, reflection, and refraction based on the Fresnel factor which describes reflection of light on the incident interface. Firstly, the index of refraction and Fresnel factor are calculated. Next, based on the incident direction and index of refraction and fresnel it is determined whether the light has total internal reflection, reflects, or refracts. If the light has total internal reflection the outgoing direction is set to the mirror of the incoming ray and BSDF and PDF are to 1.0 to model full energy reflection. If a randomly generate value is below fresnel the ray reflects in the mirror direction, the PDF is set to Fresnel value and the BSDF is set to 1.0 to indicate reflection. Lastly, if the random value is less than the Fresnel value, the ray refraction direction is calculated, the PDF is set to the refraction probability (1-Fresnel) and lastly, the BSDF is set to eta^2 by the refraction probability. This process probabilistically models light's behavior with dielectric materials. After a calculating whether an incident light ray reflects, refracts, etc. this algorithm produces the rendering appearance of transparent materials like glass or water.

funLight Generation

SceneLoader.cpp

This is a small function that generates "N" colorful QuadLights around a given model at (0, 0, 0). This allows for observance of light interactions with materials at high light source counts. It provides useful infrastructure for debugging and testing dielectric materials, RIS, and ReSTIR methods.

ReSTIR

PathTracer.cu

This code primarily implements the spatial reuse of ReSTIR due to the render scene being static. This method reuses spatial samples stored in reservoirs to resample light paths. This allows for the handling of many lights efficiently and producing noise-free results with fewer samples. To begin, first a direct lighting Reservoir is created which stores a single light sample (point, intensity, and index) as wells as the light sample's weight. Next, starting with the first frame "canonical" samples are generated per emitter present in the scene. Sample probability and weights are determined using the number of samples and a sample probability density function. One candidate is then probabilistically selected and added to a reservoir buffer. In subsequent frames, each pixel gets these reservoirs from spatial neighbors, then applies spatial importance resampling to select a new reservoir from pixel and neighbors. The newly resampled light sample has its contribution then evaluated. There are 4 key componeents to this method. First, is the generate_canonical method which creates candidate sample for the current pixel, builds a reservoir and resamples based on the BRDF and light. Next, is the evaluate_sample method which calculates the actual contribution of a light sample given a reservoir and intersection to check. Lastly, are the reservoir buffers which store neighboring pixel samples across frames in Renderer.cpp. Currently, this method of ReSTIR only considers the use of spatial reuse and not both spatial and temporal reuse. This is because the scenes rendered are static and don't change over time removing the relevance of motion vectors and frame-to-frame accumulation (camera doesn't move either in scene).

Biased ReSTIR Progress with Incorrect Weights

References

Wyman, Chris, Kettunen, Markus, Lin, Daqi, Bitterli, Benedikt, Yuksel, Cem, Jarosz, Wojciech, Kozlowski, Pawel (2023). A Gentle Introduction to ReSTIR Path Reuse in Real-Time. PDF. DOI: 10.1145/3587423.3595511. Journal: ACM SIGGRAPH 2023 Courses.

Bitterli, Benedikt, Wyman, Chris, Pharr, Matt, Shirley, Peter, Lefohn, Aaron, Jarosz, Wojciech (2020). Spatiotemporal reservoir resampling for real-time ray tracing with dynamic direct lighting. DOI: 10.1145/3386569.3392481. Journal: ACM Transactions on Graphics, Volume 39.

Dielectric BSDF (2023). PBR-Book.org.

Mitchell & Amaan ‒ Spring 2025