July 03, 2018
| Tags:

Datasets

At the moment at EGSR 2018, Walt Disney Animation Studios introduced the beginning of two large, production glorious/scale info sets for rendering analysis functions.
The details sets are readily obtainable on a brand fresh info sets page on the official Disney Animation internet space.
The foremost info region is the Cloud Recordsdata Living, which comprises a large and extremely detailed volumetric cloud info region that we frail for our “Spectral and Decomposition Tracking for Rendering Heterogeneous Volumes” SIGGRAPH 2017 paper, and the second info region is the Moana Island Scene, which is a full production scene from Moana.

Figure 1: The Moana Island Data Set, rendered using Disney's Hyperion Renderer.

Figure 2: The Cloud Data Set, rendered using Disney's Hyperion Renderer.

In this post, I’ll part some deepest solutions, observations, and notes.
The beginning of these info sets was introduced by my teammate, Ralf Habel, at EGSR recently, but this beginning has been within the works for a really very long time now, and is the fabricated from the collective effort of a immense assortment of of us across the studio.
A assortment of of us have to be highlighted: Rasmus Tamstorf spearheaded the complete effort and was instrumental in getting the resources and lawful approval wanted for the Moana Island Scene.
Heather Pritchett is the TD that did the right advanced work of extracting the Moana Island Scene out of Disney Animation’s production pipeline and converting it from proprietary info codecs into usable, industry-identical outdated info codecs.
Sean Palmer and Jonathan Garcia also helped in resurrecting the details from Moana.
Hyperion builders Ralf Habel and Peter Kutz led the anxiousness to get the Cloud Recordsdata Living authorized and released; the cloud itself was made by artists Henrik Falt and Alex Nijmeh.
On the administration side of issues, skills supervisor Rajesh Sharma and Disney Animation CTO, Prick Cannon, equipped essential toughen and encouragement.
Matt Pharr has been essential in taking part with us to get these info sets released.
Matt was extremely accommodating in serving to us get the Moana Island Scene staunch into a PBRT scene; I’ll focus on a bit extra about this later.
Intel’s Embree crew also gave essential solutions.
My position was in actuality moderately small; in conjunction with other members of the Hyperion constructing crew, I just equipped some consultation at some level of the complete direction of.

Please mask the licenses that the details sets attain with.
The Cloud Recordsdata Living is licensed under a Inventive Commons Attribution ShareAlike 3.0 Unported License; the right cloud is in step with a photograph by Kevin Udy on his Colorado Clouds Weblog, which will be licensed under the an identical Inventive Commons license.
The Moana Island Scene is licensed under a extra restrictive, customized Disney Enterprises analysis license.
That is since the Moana Island Scene is a just production scene; it was in actuality frail to assemble right frames within the final movie.
As such, the details region is being released just for pure analysis and constructing functions; it’s no longer intended for exhaust in inventive projects.
Please stick with and observe the licenses these info sets are released under; if of us stop up misusing these info sets, then it makes releasing extra info sets into the community in the end mighty extra troublesome for us.

This complete effort was sparked two years within the past at SIGGRAPH 2016, when Matt Pharr made an attraction to the industry to provide representative production-scale info sets to the analysis community.
I don’t know how repeatedly I’ve had conversations about how properly fresh recommendations or papers or technologies will scale to production cases, handiest to contain extra discussion stymied by the dearth of any just production info sets that the analysis community can take a look at against.
We determined as a studio to acknowledge Matt’s attraction, and final year at SIGGRAPH 2017, Brent Burley and Rasmus Tamstorf introduced our map to begin both the Cloud and Moana Island info sets.
It’s taken virtually a year from announcement to begin since the direction of has been advanced, and it was essential to the studio to assemble tremendous the beginning was done properly.

Undoubtedly more than likely the most supreme challenges was getting all of the details out of the production pipeline and our varied proprietary info codecs into something that the analysis community can in actuality parse and assemble exhaust of.
Matt Pharr was extraordinarily indispensable right here; over the final year, Matt has added toughen for Ptex textures and implemented the Disney Bsdf in PBRT v3.
Having Ptex and the Disney Bsdf readily obtainable in PBRT v3 made PBRT v3 the natural target for an preliminary port to a renderer instead of Hyperion, since internally all of Hyperion’s shading makes exhaust of the Disney Bsdf, and all of our texturing is performed by means of Ptex.
Our texturing also depends heavily on procedural SeExpr expressions; all of the expression-pressure texturing wanted to be baked down into Ptex for the final beginning.

Both the Cloud and Moana Island info sets are, moderately frankly, large.
The Cloud info region comprises a single OpenVDB cloud that weighs in at 2.93 GB; the details region also affords versions of the VDB file scaled all of the map in which down to half, quarter, eighth, and sixteenth scale resolutions.
The Moana Island info region is obtainable in three parts: a homely kit containing raw geometry and texture info, an animation kit containing inspiring stuff, and a PBRT kit containing a PBRT scene generated from the homely kit.
These three packages blended, uncompressed, weigh in at properly over 200 GB of disk relate; the uncompressed PBRT kit along weighs in at round 38 GB.

For the Moana Island Scene, the equipped PBRT scene requires no longer decrease than round 90 GB if RAM to render.
This many seem large for user machines, which means of it’s miles.
However, right here’s also what we mean by “production scale”; for Disney Animation, 90 GB is fully a pretty mid-fluctuate reminiscence footprint for a production render.
On a 24-core, twin-socket Intel Xeon Gold 6136 machine, the PBRT scene took me a bit over an hour and 15 minutes to render from the ‘shotCam’ digicam.
Hyperion renders the scene sooner, but I’d caution against the exhaust of this info region to originate performance shootouts between varied renders.
I’m particular that within a brief duration of time, alive to members of the rendering community will stop up porting this scene to Renderman and Arnold and Vray and Cycles and each other production renderer available within the market, that may perhaps be very cool!
However take be conscious of, this info region was authored very namely round Hyperion’s varied capabilities and constraints, which naturally will be very varied from how one may perhaps creator a posh info region for other renderers.
Every renderer works a bit in another map, so the most optimal map to creator an info region for every renderer will be a bit varied; this info region is rarely any exception.
So while you’re going to love to evaluate renderers the exhaust of this info region, assemble tremendous you respect the many ways how the approach this info region is structured impacts the performance of whatever renderers you are evaluating.

Let’s snarl, Hyperion subdivides/tessellates/displaces everything to as shut to sub-poly-per-pixel as it may perhaps perhaps perhaps get while nonetheless fitting within computational resources.
This methodology our scenes are on the complete very heavily subdivided and tessellated.
However, the PBRT version of the scene doesn’t attain with any subdivision; which means, silhouettes within the next comparison photos don’t fully match in some areas.
In a similar map, PBRT’s lights and lighting model fluctuate from Hyperion’s, and Hyperion has varied inventive controls that are recurring to Hyperion, which methodology the renders produced by PBRT versus Hyperion fluctuate in many ways:

Figure 3a: 'shotCam' camera angle, rendered using Disney's Hyperion Renderer.

Figure 3b: 'shotCam' camera angle, rendered using PBRT v3.

Figure 4a: 'beachCam' camera angle, rendered using Disney's Hyperion Renderer.

Figure 4b: 'beachCam' camera angle, rendered using PBRT v3.

Figure 5a: 'dunesACam' camera angle, rendered using Disney's Hyperion Renderer.

Figure 5b: 'dunesACam' camera angle, rendered using PBRT v3. Some of the plants are in slightly different locations than the Hyperion render; this was just a small change that happened in data conversion to the PBRT scene.

Figure 6a: 'flowersCam' camera angle, rendered using Disney's Hyperion Renderer.

Figure 6b: 'flowersCam' camera angle, rendered using PBRT v3. Note that the silhouette of the flowers is different compared to the Hyperion render because the Hyperion render subdivides the flowers, whereas the PBRT render displays the base cage.

Figure 7a: 'grassCam' camera angle, rendered using Disney's Hyperion Renderer.

Figure 7b: 'grassCam' camera angle, rendered using PBRT v3. The sand dune in the background looks particularly different from the Hyperion render due to subdivision and displacement.

Figure 8a: 'palmsCam' camera angle, rendered using Disney's Hyperion Renderer.

Figure 8b: 'palmsCam' camera angle, rendered using PBRT v3. The palm leaves look especially different due to differences in artistic lighting shaping and curve shading differences. Most notably, the look in Hyperion depends heavily on attributes that vary along the length of the curve, which is something PBRT doesn't support yet. Some more work is needed here to get the palm leaves to look more similar between the two renders.

Figure 9a: 'rootsCam' camera angle, rendered using Disney's Hyperion Renderer.

Figure 9b: 'rootsCam' camera angle, rendered using PBRT v3. Again, the significant difference in appearance in the rocks is probably just due to subdivision/tesselation/displacement.

One more example of a predominant distinction between the Hyperion renders and the PBRT renders is within the water, which Hyperion renders the exhaust of photon mapping to get the caustics.
The equipped PBRT scenes exhaust unidirectional pathtracing for everything in conjunction with the water, hence the very varied caustics.
In a similar map, the palm bushes within the ‘palmsCam’ digicam attitude glimpse very varied between PBRT and Hyperion which means of Hyperion’s lighting controls are very varied from PBRT; Hyperion’s lights contain varied inventive controls for customized shaping and whatnot, which aren’t necessarily fully physical.
Moreover, the palm leaves are modeled the exhaust of curves, and the shading is dependent on varied colors and attributes along the scale and width of the curve, which PBRT doesn’t toughen but (getting the palm leaves is fully the tip precedence for if extra resources are freed up to beef up the details region beginning).
These distinction between renderers don’t necessarily mean that one renderer is extra healthy than the opposite; they simply mean that the renderers are varied.
This will be just for any pair of renderers that one needs to evaluate.

The Cloud Recordsdata Living contains an example render from Hyperion, which implements our Spectral and Decomposition Tracking paper in its volumetric rendering machine to efficiently render the cloud with thousands of bounces.
This render comprises no post-processing; what you gaze within the equipped image is strictly what Hyperion outputs.
The VDB file expresses the cloud as a discipline of heterogeneous densities.
Moreover equipped is an example Mitsuba scene, renderable the exhaust of the Mitsuba-VDB plugin that may perhaps be found on Github.
Please consult the README file for some modifications in Mitsuba that are essential to render the cloud.
Moreover, please mask that the Mitsuba example will take dangle of an especially very long time to render, since Mitsuba isn’t in actuality intended to render high-albedo heterogeneous volumes.
With upright acceleration structures and algorithms, rendering the cloud handiest takes us a brief time the exhaust of Hyperion, and wants to be equally fleet in any current production renderer.

One may perhaps surprise just why production info sets in identical outdated are so large.
That is a mesmerizing inquire; the brief acknowledge across the industry on the complete boils all of the map in which down to “artist time is dearer and precious than computer hardware”.
We may perhaps get these scenes to suit into mighty smaller footprints if we were inspiring to assemble our artists use plenty of time aggressively optimizing property and scenes and whatnot so as that we may perhaps fit these scenes into smaller disk, reminiscence, and compute footprints.
However, this isn’t in actuality all the time a ethical exhaust of artist time; computer hardware is low-fee when put next with losing artist time, which often will be better spent in other areas making the movie better.
Throwing extra reminiscence and whatnot at large info sets will be simply extra scalable than the exhaust of extra artist resources, moderately talking.

Both info sets attain with detailed README documents; the Moana Island Scene’s documentation in particular in all fairness extensive and comprises a predominant quantity of details about how property are authored and structured at Disney Animation, and the map in which renders are lit, art-directed, and assembled at Disney Animation.
I extremely counsel learning all of the documentation carefully while you intend on working with these info sets, or just while you are on the complete extraordinary about how production scenes are constructed at Disney Animation.

Personally, I’m very mighty having a glimpse forward to seeing what the rendering community (and the wider computer graphics community at large) does with these info sets!
I’m namely mad to glimpse what the realtime world will have the selection to originate with this info; seeing the Moana Island Scene in its full glory in Unreal Engine 4 or Team spirit may perhaps be something indeed, and I judge these info sets ought to provide an out of this world anxiousness to analyze into light transport and ray tracing tempo as properly.
Whenever you happen to originate attention-grabbing issues with these info sets, please write to us at the email addresses within the equipped README recordsdata!

Moreover, Matt Pharr has written on his blog about how the Moana Island Scene has extra driven the development of PBRT v3.
I extremely counsel giving Matt’s blog a learn!