The November 7 arrival of Disney’s Big Hero 6 is as much a supercomputing triumph as it is an animated feast. It’s the coming out party for Hyperion—a cutting-edge light rendering software shaped by both Disney artists and engineers working in concert for two years.
The system hauled 200 million computing hours, spawned a number of companion programs, and enabled animators to imbue the film’s fantastical settings, elements, and characters with a realism and dimension that would have otherwise been impossible.
“It allows us to put more on the screen with the same number of artists, creating a richer world that better supports the story,” says Disney Animation chief technology officer Andy Hendrikson.
The Japanese influenced film—directed by Don Hall and Chris Williams from an obscure Marvel property—chronicles of a group of social misfits and ingenuous robot as they attempt to save their city of San Fransokyo (a mash-up of San Francisco and Tokyo) from an evil technologist. Hyperion was also used for the Disney short, Feast, that will precede BH6 screenings during its theatrical run.
Hyperion tracks how light rays bounce off multiple objects in an environment before they enter your eyes. When a ray hits an object with a diffuse surface–say, something with a satin sheen–it scatters in many directions. Each of the scattered rays continue bouncing off other objects until they all ultimately lose energy and dissipate. The result is softer, diffused light, more nuance and shadows, creating a more realistic look.
But getting there is a unique kind of hell. Tracking those hundreds, possibly millions, of rays in all directions would overload computers’ random access memory (RAM), which stores quick-retrieval data during operation.
“Before, we’d been hampered by the level of computation needed at any one time, the amount of memory needed to store it, and the cost,” says Hendrickson. “Ten years ago, the level of investment would have been greater than the cost of two to three films. Hyperion allowed a 100-fold increase in image complexity for a fraction of the [undisclosed] cost.”
Hyperion solves the storage and computational problems by condensing the number of algorithms to be calculated in a given amount of time, while increasing the number of light rays traced per second. The software winnows the light scatter into six main directions. As each ray hits another object, the program divides that scatter into another six directions, and so forth, until the light rays disappear. Each time a ray hits and refracts off an object, it’s called a bounce. Hyperion carried out most of its scattered light vectors to 10 bounces.
From there, Hyperion applies another set of algorithms, called filtering passes, that interpolate the additional rays that, in reality, would have occurred between the six directions, giving the appearance of infinite light bounces and a more nuanced glow. Without Hyperion, animators would have rendered this manually, which, given the image complexity in this movie, would have been impossible in the allotted production time.
Ten bounces was also crucial for the translucent look of one of the film’s main characters, a soft robot named Baymax. “If we kept it to two bounces, he ended up looking like hard plastic,” says Hendrickson. “But 10 light bounces taking place inside his body, before the light finally emerged, gave him his translucence.”
Disney Animation increased its processing power by connecting four rendering farms–three in Los Angeles and one in San Francisco–into a giant supercomputer of nearly 4600 computers running 55,000 cores–microchips containing CPU processors. (By comparison, Frozen used 26,000 cores). The division’s previously developed automated management system, Coda, guided the information flowing between the four rendering farms, enabling it to process 400,000 rendering jobs overnight for the next morning.
Complementing Hyperion, Disney coders created additional automation software to expedite and enhance other areas of rendering. One process extrapolated algorithms describing San Francisco and Tokyo architectures to create city buildings. Another, code named Denizen, created 16,000 unique city occupants by recombining portions of thousands of different base characters in a variety of textures, fabrics, and colors. (This is different than crowd simulation software that simply replicates 10 body types.) Another in-house program, Bonzai, automated the creation of the city’s 250,000 trees. The process involved interspersing the building and rendering layers of trees, people, and architecture.
“It was easy enough to use so the animators could make avatars of themselves and place them in the city as extras,” says Hendrickson. “So a lot of folks in the studio are also in the movie.”
Still another custom process directed the movement of the microbots, spindle-shaped micro-computers—some 20 million onscreen in a given shot—that both swarmed in tandem and hurdled over one another in circuit-like waves, eventually organizing into structures. (This is more algorithmically complex than the flowing motion of the iconic birth of Sandman scene in Spider-Man 3.)
Disney Animation embarked upon the BH6 story four years ago, but only started working on Hyperion two years ago, gleaning constant input from a user base of 400 artists and programmers. Sometimes, the artists would ask the programmers to attempt a tool enabling them to realize a certain look; other times, the programmers achieved a simulation that gave the artists new ideas for visualizing the film.
“It’s pretty aggressive building something up that fast, but we were trying to make a world that supported an awesome story,” says Hendrickson, who hopes the collective software will facilitate a film a year. “We were exploring the art and algorithms simultaneously, riffing off each other, building the renderer at the same time they were using it to make art. But without these methods, we couldn’t handle the complexity of this world. We consider our in-house software to always be in beta, because we want it to always be evolving.”
P.S. When you see the film, stay till the very end, past the credits.