In Season 2 of the HBO sci-fi/drama The Last of Us, Joel and Ellie are drawn into conflict with each other and a world that's become even more dangerous and unpredictable. Pedro Pascal returns as Joel, a hardened survivor who’s been hired to smuggle 14-year-old Ellie (Bella Ramsey), out of an oppressive quarantine zone. There are several new additions to the cast of Season 2, including Kaitlyn Dever, Isabela Merced, Young Mazino, Ariela Barer, Tati Gabrielle, Spencer Lord, Danny Ramirez, Jeffrey Wright and Catherine O’Hara.
Wētā FX’s VFX supervisor Nick Epstein calls attention to a number of challenging VFX sequences that the studio worked on for Season 2.
“The Jackson siege in Episode 2 was an interesting challenge, as the scope was so huge,” Epstein shares. “In terms of the single biggest problem it presented, I would say it was ensuring the CG infected horde looked realistic and sat in the photography with their real-world counterparts. As with any big problem, we began by breaking it down into more manageable chunks.”
The VFX team had the benefit of readily available stunt actors in prosthetics, so they started by selecting which of them best fit the vibe (creator/writer) Craig Mazin was going for with production VFX supervisor Alex Wang. They were scanned on-set and Weta’s asset departments then got to work building 1:1 digital versions.
“The stunt horde was very 'level 0' in terms of infection progression, so we did a lot of concepting to figure out how much additional infection would work to add variety to the horde as a whole, particularly when it would be seen as a combination of practical and CG infected,” he shares. “The selects were level 1 to level 2, with more cordyceps protrusions, more textural infection, and continuation of infection where the prosthetics stopped around the eyes, nose, mouth and ear sockets mainly.”
According to Epstein, the team had just under 30 base infected builds, which looked great individually, but there was still a need for hundreds of infected characters, so variation presented a further challenge.
“Through initial crowd tests we discovered primarily hair and clothing were most noticeable in terms of repetition,” he reveals. “So, we decided to see if we could modularize all of the clothing that we had built from the scans, which was not straightforward, as they all had (sometimes drastically) varying proportions. Thankfully our creature lead came up with a dynamic refitting setup that allowed any infected to wear any piece of clothing, and a similar setup for inheriting hair grooms, so we generated close to 400 different collections defining all of the geometric possibilities, knowing (and honestly initially praying) that during simulation everything would align.”
The shading of the clothing also needed to be varied.
“We started authoring color variants with the usual controls for offsetting colors, controlling snow and dirt levels, but this didn't amount to meaningful variation when looking at the full horde together,” Epstein notes. “Between our look-dev and CG supervisors, we arrived at a solution that allowed our lighting artists to procedurally texture the horde's clothing directly in the context of a shot. It was maybe a week later that we started getting notes along the lines of, 'We need a red checked shirt over there, and could we get some yellows and whites over here,’ which would have been quite problematic to address without this system.”
Motion also presented its own challenges, particularly for the performance of actions that stunt or motion capture actors couldn’t perform, such as slamming into a wall or the ground, being trampled upon, or being caught in an explosion.
“We needed to incorporate ragdoll dynamics into our approach, which again, was not straightforward, as we also needed to be able to simulate hair and cloth on top of the ragdoll simulations, which often did not obey anatomy,” he explains. “And on top of cloth and hair solver constraints, we also had lots of shots where the infected needed to be on fire – and on fire in quite a particular way – meaning all of the input motion and ragdoll simulations needed to be within a set of bounds that we could tune our pyro solver to.”
Another realism factor that needed attention was how the horde emoted. Every infected individual had a full FACS-based facial rig, which allowed for both randomized 'contorted screaming,' but also specific event-driven performance, like being shot, attacking and biting.
“There was definitely a degree of, ‘This better work,’ as we had to action all of this development in a lot of different areas simultaneously, so it was quite nerve wracking as we approached the first looks at our horde shots. The need here was basically for a crowd of hero characters and we ended up with the highest number of fully hair and cloth simulated bipeds we've ever dealt with. One of our creature artists was responsible for 61 year’s worth of simulation!”