University of Pennsylvania, CIS 565: GPU Programming and Architecture, Project 4
- Zijing Peng
- Tested on: Windows 22, i7-8750H@ 2.22GHz 16GB, NVIDIA GeForce GTX 1060 (laptop)
In this project, I implemented a pathtracing denoiser that uses geometry buffers (G-buffers) to guide a smoothing filter, base on the paper "Edge-Avoiding A-Trous Wavelet Transform for fast Global Illumination Filtering" by Dammertz, Sewtz, Hanika, and Lensch.
The following figure illustrates the basic steps of our algorithm. The input is the path-traced (noisy) image, along with a normal buffer and a position buffer. The algorithm use Gaussian filter with increasing step width for multiple iterations to get a denoised result.
The following figures show the renders with different denoising iterations. Compared with the PT reference, the result of 5 iterations shows best visual effect among those candidates, it is relatively "smooth", and the shadow looks good.
| Input | PT Reference |
|---|---|
 |
 |
| 2 iterations | 5 iterations |
 |
 |
| 10 iterations | 100 iterations |
 |
 |
According to the profiling below, he denoising time increases approximately linearly with the number of times.
The material type does not affect the denoising time, since this algorithm is "post-processing".
Based on path tracing algorithm, a smaller light source means more samples needed to get a smooth result. Thus, if we take 10 iterations for path tracing (like we do before), the result is much nosier. More denoising iterations you'll add to bring a smooth result. However, although it is smooth, it is kind of dark. Another choice is to do the path tracing for more iterations before doing denoiser (see the last figure).
| Large Light + PT 10 iterations | Small Light + PT 10 iterations | Small Light + PT 5000 iterations |
|---|---|---|
|
 |
 |
| Denoising 5 iterations + 10 PT | Denoising 100 iterations + 10 PT | Denoising 5 iterations + 5000 PT |
 |
 |
 |