Oz: The Great and Volumetric Magnus Wrenninge Sony Pictures Imageworks Chris Kulla Sony Pictures Imageworks Viktor Lundqvist Sony Pictures Imageworks Figure 1: Renders from Oz: The Great and Powerful. The left image shows a water simulation with volumetric spray and foam elements. The right image shows the effect of the multiple scattering approximation. c 2013 Disney Enterprises Inc. For Oz: The Great and Powerful, Imageworks provided over 1200 shots, 400 of which included substantial volumetric elements. We discuss the challenges of various sequences and optimizations to our renderer performed during the course of production. Path traced volumes in production Starting on The Amazing Spiderman and Men In Black 3, Imageworks’ in-house version of Arnold has supported path tracing of volumes alongside surfaces. The pipeline is built around the Field3D file format, and volume primitives support all standard light sources, including area lights and textured skydomes. We maintain reasonable render times with the importance sampling techniques detailed in [Kulla and Fajardo 2012]. To lighters, volumes were seen as just another scene graph location in Katana, with a geometric proxy representation for fast previewing and layout. Named attributes of each field automati- cally bind to similarly named shader parameters, so that TDs can write custom shaders easily. Under the hood, the Field3D primitive supported arbitrary numbers of fields in each .f3d file, and had full support for dense, sparse and MAC fields of mixed bit depths. Multiple scattering approximation A large portion of the volu- metric work on Oz consisted of fog, mist, steam and clouds, all of which are high albedo media. Inspired by the work of Bouthors [2011] on cloud rendering, where multiple shadow maps were used to approximate different orders of multiple scattering, we devel- oped an approximation for path tracing that incurs little render time overhead, adds little extra noise, and can be importance sampled. The main idea is to artificially lower the extinction coefficient σt along the shadow ray to let more light reach the shaded point. But rather than use a fixed scaling factor, we use a summation over several scales. We also adjust the local phase function eccentricity g and local scattering coefficient σs such that the total contribution of light at a given point is: L = N-1 X i=0 Li (1) Li = σsb i L light (ωi )p ( ωi ,ωo,c i g0 ) e -a i R t 0 σ t (s)ds (2) where N (“octaves”), a (“attenuation”), b (“contribution”), and c (“eccentricity attenuation”) are user-tunable parameters. De- faults of N = 8 and a = b = c =1/2 are a good starting point for all the cases we tried. Since the change in eccentricity in the phase function makes the choice of sampling direction more complex, we stochastically choose a single octave with probability of b i . This allows keeping the ray budget constant and renders with little extra noise. These parameters gave artists simple con- trol over the multiple scattering effect, without having to tweak scattering lengths manually. This approximation can be used on camera-shadow rays when only simulating direct lighting, or on camera-indirect-shadow rays if already simulating one bounce. The latter option is more expensive but more accurately captures the diffusion effect. Shading system Volumes were also integrated into Arnold’s OSL-based shading system, which allowed TDs to build custom shaders for volumes in Katana, just as they would for surface shaders. In order to provide a standardized interface to the lighting pipeline, a “root” illumination shader was provided to the TDs, into which arbitrary shader nodes were hooked. Several sequences relied heavily on custom shaders, for effects such as fire, explosions, and ash clouds, and also for procedural noise banks used to quickly layout generic fog. Efficient traversal of frustum buffers For the case of sparse uniform volumes, we limit ray marching to occupied blocks by using Bresenham traversal. However, for frustum buffers there is no guarantee that an arbitrary ray remains straight in voxel space, as the perspective division may bend a ray into a curved path. Rather that attempt some form of curved ray marching, we observe that the projection mapping does not bend axis-aligned planes and build a kd-tree over the sparse blocks in voxel space. During ray intersections tests, the planes are transformed into world space which avoids testing against curved rays. To support motion blurred mappings and reduce the size of the tree we perform the plane transform on the fly during traversal. Since the planes are axis-aligned, this is a very simple operation on the perspective matrix entries. We pay special attention to numerical robustness to ensure the continuity of the detected occupied intervals. This sped up the integration of high resolution buffers greatly, saving the overhead of shader calls in regions that are known to be empty. As our method makes no assumption about ray direction, it applies equally well to shadow rays and other non-view-aligned rays. References Kulla, C., and Fajardo, M. 2012. Importance sampling techniques for path tracing in participating media. Computer Graphics Forum 31, 4, 1519–1528. Wrenninge, M., Bin Zafar, N., Harding, O., Graham, G., Tessendorf, J., Grant, V., Clinton, A., and Bouthors, A. 2011. Production volume rendering 2: Systems. ACM SIGGRAPH 2011 Courses.
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Oz: The Great and Volumetric
Magnus WrenningeSony Pictures Imageworks
Chris KullaSony Pictures Imageworks
Viktor LundqvistSony Pictures Imageworks
Figure 1: Renders from Oz: The Great and Powerful. The left image shows a water simulation with volumetric spray and foamelements. The right image shows the effect of the multiple scattering approximation. c©2013 Disney Enterprises Inc.
For Oz: The Great and Powerful, Imageworks provided over1200 shots, 400 of which included substantial volumetric elements.We discuss the challenges of various sequences and optimizationsto our renderer performed during the course of production.
Path traced volumes in production Starting on The AmazingSpiderman and Men In Black 3, Imageworks’ in-house version ofArnold has supported path tracing of volumes alongside surfaces.The pipeline is built around the Field3D file format, and volumeprimitives support all standard light sources, including area lightsand textured skydomes. We maintain reasonable render timeswith the importance sampling techniques detailed in [Kulla andFajardo 2012].
To lighters, volumes were seen as just another scene graphlocation in Katana, with a geometric proxy representation for fastpreviewing and layout. Named attributes of each field automati-cally bind to similarly named shader parameters, so that TDscan write custom shaders easily. Under the hood, the Field3Dprimitive supported arbitrary numbers of fields in each .f3d file,and had full support for dense, sparse and MAC fields of mixedbit depths.
Multiple scattering approximation A large portion of the volu-metric work on Oz consisted of fog, mist, steam and clouds, all ofwhich are high albedo media. Inspired by the work of Bouthors[2011] on cloud rendering, where multiple shadow maps were usedto approximate different orders of multiple scattering, we devel-oped an approximation for path tracing that incurs little rendertime overhead, adds little extra noise, and can be importancesampled. The main idea is to artificially lower the extinctioncoefficient σt along the shadow ray to let more light reach theshaded point. But rather than use a fixed scaling factor, we usea summation over several scales. We also adjust the local phasefunction eccentricity g and local scattering coefficient σs suchthat the total contribution of light at a given point is:
L =
N−1∑i=0
Li (1)
Li = σsbiLlight(ωi)p
(ωi, ωo, c
ig0)e−ai∫ t
0σt(s)ds
(2)
where N (“octaves”), a (“attenuation”), b (“contribution”), andc (“eccentricity attenuation”) are user-tunable parameters. De-faults of N = 8 and a = b = c = 1/2 are a good starting pointfor all the cases we tried. Since the change in eccentricity inthe phase function makes the choice of sampling direction morecomplex, we stochastically choose a single octave with probabilityof bi. This allows keeping the ray budget constant and renders
with little extra noise. These parameters gave artists simple con-trol over the multiple scattering effect, without having to tweakscattering lengths manually. This approximation can be used oncamera-shadow rays when only simulating direct lighting, or oncamera-indirect-shadow rays if already simulating one bounce.The latter option is more expensive but more accurately capturesthe diffusion effect.
Shading system Volumes were also integrated into Arnold’sOSL-based shading system, which allowed TDs to build customshaders for volumes in Katana, just as they would for surfaceshaders. In order to provide a standardized interface to thelighting pipeline, a “root” illumination shader was provided tothe TDs, into which arbitrary shader nodes were hooked. Severalsequences relied heavily on custom shaders, for effects such asfire, explosions, and ash clouds, and also for procedural noisebanks used to quickly layout generic fog.
Efficient traversal of frustum buffers For the case of sparseuniform volumes, we limit ray marching to occupied blocks byusing Bresenham traversal. However, for frustum buffers thereis no guarantee that an arbitrary ray remains straight in voxelspace, as the perspective division may bend a ray into a curvedpath. Rather that attempt some form of curved ray marching, weobserve that the projection mapping does not bend axis-alignedplanes and build a kd-tree over the sparse blocks in voxel space.During ray intersections tests, the planes are transformed intoworld space which avoids testing against curved rays. To supportmotion blurred mappings and reduce the size of the tree weperform the plane transform on the fly during traversal. Sincethe planes are axis-aligned, this is a very simple operation on theperspective matrix entries. We pay special attention to numericalrobustness to ensure the continuity of the detected occupiedintervals. This sped up the integration of high resolution buffersgreatly, saving the overhead of shader calls in regions that areknown to be empty. As our method makes no assumption aboutray direction, it applies equally well to shadow rays and othernon-view-aligned rays.
References
Kulla, C., and Fajardo, M. 2012. Importance samplingtechniques for path tracing in participating media. ComputerGraphics Forum 31, 4, 1519–1528.
Wrenninge, M., Bin Zafar, N., Harding, O., Graham, G.,Tessendorf, J., Grant, V., Clinton, A., and Bouthors,A. 2011. Production volume rendering 2: Systems. ACMSIGGRAPH 2011 Courses.
Related Documents
