Another real-time rendering test with Brigade, set in the Urban Sprawl scene.
- 800x480 resolution
- 5 spp
- 2x GTX 580
- because there is very little noise, blurring (frame averaging) is turned off
UPDATE: A reader of this blog sent me a highly detailed mesh of a procedural mountain landscape (500k traingles), created with World Machine, which rendered extremely fast with Brigade. The image below took less than half a second to render (the lighting isn't great, but that's not the point of this test :)
UPDATE 2: some pictures of an upcoming demo (idea by Jeroen van Schijndel):
the refraction index of the window glass is probably too much, should be tweaked
There's a new trailer of "It's about time", a real-time path traced game running on Brigade, created by a team of eight IGAD students (International Game Architecture and Design) from the NHTV, the Netherlands:
Another test with the ray traced AO kernel in Brigade. This kernel holds the middle between distributed ray tracing [Cook] and full path tracing [Kajiya] and provides soft shadows, refraction, reflection (glossy and perfectly specular), depth of field and ambient occlusion with adjustable ray length, all raytraced in real-time and without screen space limitations or other hacks.
The best thing about this kernel is that the image converges so insanely fast that I've decided to tone down the frame averaging (blur) almost to zero. The little noise that still remains is mostly visible around shadow edges and on reflecting surfaces, but it's almost negligible compared to the amount of noise when using the path tracing kernel. Brigade is now extremely close to real-time noise-free rendering and this is still without any fancy noise filters.
If the current rate of development continues, you might expect to see the Brigade engine at E3 next year (disclaimer: kidding)
Real-time rendered test at 800x480 resolution with 6 samples per pixel on 2x GTX 580:
UPDATE: According to this article, Unreal Engine 4 uses a form of GPU based real-time ray tracing for global illumination called voxel cone tracing in a sparse voxel octree (SVOGI). The technique was pioneered by Cyril Crassin and was presented at Siggraph last year (sparse voxel octrees were all the hype in 2008 thanks to Jon Olick's Siggraph demo, which was researched further independently by Cyril Crassin and Samuli Laine). The future of real-time ray tracing on the GPU looks bright with one of the largest game companies starting to use it on a large scale :)
UPDATE 2: A reader of this blog just sent me a screenshot of an awesome stress test with the public Brigade version: a giant cube containing 729k (90x90x90) tiny cubes, a total of 8.75 million triangles, consuming ~ 2 GB of video memory. Thanks Nicholas!
Testing Brigade's latest gamma correction update with the Sponza scene (the Bugtoad was modeled by Son Kim):
Recommended to watch in 720p:
Today, I've also stumbled upon a interesting paper entitled "Toward Evaluating Progressive Rendering Methods in Appearance Design Tasks" by Ou, Karlik, Krivanek and Pellacini. The authors tested 4 progressive rendering methods (random path tracing, quasirandom path tracing, progressive photon mapping and virtual point lights rendering) and investigated how people subjectively perceive and enjoy the progressive updating of the image.
Quoting the abstract (the full paper can be found here):
"Progressive rendering is becoming a popular alternative to precomputation approaches for appearance design tasks. Images created by different progressive algorithms exhibit various kinds of visual artifacts at the early stages of computation. We present a user study that investigates the effects of these artifacts on user performance in appearance design tasks. Specifically, we ask both novice and expert subjects to perform lighting and material editing tasks with the following algorithms: random path tracing, quasi-random path tracing, progressive photon mapping, and virtual point light (VPL) rendering. Data collected from the experiments suggest thatpath tracing is strongly preferred to progressive photon mapping and VPL rendering by both experts and novices. There is no indication that quasi-random path tracing is systematically preferred to random path tracing or vice-versa; the same holds between progressive photon mapping and VPL rendering. Interestingly, we did not observe any significant difference in user workflow for the different algorithms. As can be expected, experts are faster and more accurate than novices, but surprisingly both groups have similar subjective preferences and workflow."
During the first frame updates, progressive photon mapping exhibits low frequency noise in the form of ugly splotches, while VPL rendering suffers from banding artefacts. The noise produced by path tracing on the other hand is much more easy on the eyes showing that the human visual system is more forgiving for Monte Carlo noise.