Having just completed work on the second season of The Last of Us as VFX Supervisor at Weta FX, Nick Epstein brings an impressive VFX pedigree. Over the past 20 years, he has worked across leading studios such as Weta FX, Framestore, Digital Domain, and Rising Sun Pictures, contributing to Avatar: The Way of Water, The Golden Compass, Children of Men, and Godzilla x Kong.
After talking about his work on The Last of Us Season 1 back in 2023, and then working on Kingdom of the Planet of the Apes, Dennis Yoo is now back with us to break down the challenges of the show’s highly anticipated second season.
How did you get involved on this show?
Nick Epstein (NE) – I worked with Alex many years ago at Digital Domain and we kept in touch. My family and I had planned a move to Weta FX’s Vancouver hub for a while and thankfully everything aligned for the start of season two of The Last of Us!
Dennis Yoo (DY) – For season one, when I heard we might be working on The Last of Us, I was overly excited, as I’m a fan of the games. I made it known that I wanted to supervise Animation on the show, and I’m lucky enough to have been fortunate to have worked on both seasons so far.
Nick, what was your feeling to be part of this iconic universe?
NE // Nervously aware of the lofty expectations after the incredible season one! But also really excited as personally the second game impacted me a lot, I thought it was a masterpiece, so definitely a dream come true to be able to contribute to the series.
How was the collaboration with the showrunner and VFX Supervisor Alex Wang?
NE // I worked closely with Alex throughout. He had me on set for the Jackson siege part of the shoot, which meant we could problem-solve and make decisions together early on and then have a shorthand later in post. There was a definite commitment to excellence that came from the top down which, while this was certainly a hard show, led to the best VFX work I’ve been part of.
DY // Our main point of contact was through Alex; there was a definite camaraderie when coming up with solutions to problems. Alex would even find useful references/clips and imagery that he shared with us when pertinent. Craig and Neil made recordings for us on rare occasions when needed. They were expecting excellence which reflected in their notes, and we were game to achieve it.
How were the sequences made by Weta FX?
NE // We were responsible for nearly 400 shots across the first five episodes, with the bulk of our work being episode two. We handled the horde emerging from under the snow and chasing Abby, through the whole siege of Jackson, save for the Tommy/Bloater alleyway encounter which was ILP. We did some Jackson exterior environment shots in episodes 1 and 3, the subway sequence in episode 4 and the warehouse Stalkers in episode 5.
Outside of previs, post visual effects work usually begins once we receive the first edit of a sequence. We’re assigned key “test” shots that represent the various types of work in the sequence, and that’s where we begin to work our magic—exploring cameras and track, integrating assets and creatures, and applying animation, FX, lighting, and compositing in context. Much of this early work is exploratory; some of it ultimately isn’t used in the final cut, but it’s all part of the process of discovering what the sequence really needs. These test shots help us refine the creative and technical direction for the rest of post and guide each specialized department moving forward.
You created previs for the Jackson siege in episode two. How closely did that previs translate to what was actually shot on set, and what kind of collaboration did it require with the showrunners and directors?
NE // Mike Cozens supervised the Jackson Siege previs and I was totally blown away by the job he and his team did. He visited the set a few times to get a tactile feel for the place and went through a number of revisions with Mark Mylod, the episode director, Alex, and Craig in a pretty tight timeframe. It was pretty cool to be on set and see the previs being referred to as setups were being done, everyone was excited by it and much of the action was shot as designed.
DY // I was on another project during that time unfortunately, but I was involved in the initial previs motion capture based off of the script from Craig, and some previs on the tail end of the Jacksonville shoot. In general, we built a virtual model of the Jackson set during construction, using live set plans for previs. We set up shots during our capture sessions, using CG cameras that were grounded with realistic motion. This gave the onset team an early look at the feel of the shots and allowed real-time adjustments as the set was built. It was a very interesting and rewarding process to see the previs work come to life as real filmed plates in some cases.
Episode two features over 1,000 Infected emerging from beneath the snow. How did you manage the animation of such a massive crowd while maintaining unique, believable motion for each creature?
DY – While there were several shots in episode two that were highly complex, that shot was arguably the most challenging on multiple levels. Our first major hurdle was getting alignment on the shot’s intent—what exactly was the story we were trying to tell? Helping Craig visualize this was both exciting and daunting. After several months and numerous previs iterations, we all began to grasp the sheer scope of this monumental shot.
The next big question was how to actually execute it.
We started by getting buy-off on the shape of the crevasse, which was established through CG modeling. Fortunately, we had a specialized motion artist on our team, Andre Castelao, whose expertise lies in ragdoll physics simulations and motion editing. He used simulation to fill the crevasse with Infected bodies, forming the base layer of the mass.
On top of that, we placed mocap performances of Infected climbing and crawling. Once we were happy with how the base layer moved, we added another ragdoll simulation pass to allow them to push through the body barricade—creating believable gaps to climb through.
From there, multiple animators and motion artists meticulously keyframed and motion-edited the interactions of Infected grabbing and climbing on top of one another—layer by layer. All the while, we continued iterating with cloth simulations and FX-driven snow passes. Finally, the shot was handed off to lighting and comp to complete this monstrous sequence.
Every Infected was created to hero level, with facial motion shapes and simulated hair and clothing. Why was it important to build each Infected at such a high fidelity, and how did that impact your animation process?
NE // We knew from the shoot we had a lot of hero shots where our Infected were going to be seen full screen, often in between or next to stunt actors, so essentially, we were building a crowd of hero characters. We also had a lot of mid-distance shots where lack of cloth and hair simulation would have been very evident. I think this was the highest number of fully simulated characters we’ve ever done.
DY // The benefit of having the Infected built with higher fidelity is we didn’t have to worry about which Infected ran close to camera, which is normally the case if we build lower 1k character assets that don’t hold up as hero. It ultimately meant we could increase variety and have less repetition of which ones were seen up close.
In episode five, we see the Stalkers in a tense warehouse sequence. What specific animation or behavioral traits did you design to distinguish them from the other Infected types?
DY // The way they moved was already established by a specific performer Craig liked. We were given extensive reference footage of her, as she had a very peculiar and difficult-to-replicate movement style, fast, twitchy motions combined with an odd crawling lift that required significant body strength and conditioning. We used her performance as reference and keyframe-animated over mocap, since it was extremely difficult to find anyone else who could move like her.
The Bloater was entirely redesigned for season two. Can you walk us through the creature’s development, from concept to final shots, and the specific challenges introduced by the snow and frost-covered environment?
NE // The Bloater needed to be a ‘Holy shit!’ moment. Jackson was already under siege and then things get even worse when Tommy hears a roar out there in the blizzard, then we see him. So we made him bigger than his season one counterpart, I think we landed on 10% – more imposing and powerful, with armor-like cordyceps details including a distinctive crown, which our asset supervisor Pascal Raimbault sculpted from photogrammetry of chicken-of-the-woods fungus. We covered him in a geometric frost layer and scattered ice buildup in appropriate places – he had to look like he had just charged through a snowy tundra along with the rest of the horde. The lighting conditions certainly factored into his development. Diffused bright daylight gives you nowhere to hide, and dropping our already excellent season one Bloater into the season two environment just confirmed how unforgiving this scenario was going to be – I think during the rebuild we at least doubled our texture resolution and especially with the frosting, we carried a lot more geometry than we had for season one.
The avalanche-buried Infected burst through the snow in a chaotic sequence. What were the main challenges in terms of FX for snow interaction, destruction, and gore?
NE // That initial reveal of the horde was one of the hardest challenges we had. We did a lot of revisions on the Clicker head eerily descending into the snow. That motion was actually arrived at via a fluid sim, which then in turn influenced the snow simulation. As tricky as that was, it was just a warmup for the ‘disturbed ant hive’ of Infected that came right after. That shot was a sieve of (ragdoll dynamic driven) frozen bodies descending into the snow as frenzied live Infected clambered up through them. Everyone had fully simmed cloth and hair, and snow simulation specialist Igor Bondar ended up with 21 separate sims all with additional secondaries, interacting with the Infected.
You also created CG burning buildings and atmospheric effects to enhance the Jackson attack. How did you balance practical effects with CG augmentation, especially for fire and smoke?
NE // I really love the grounded approach Alex, and The Last Of Us generally, takes to VFX. There’s always a plate, even if it’s ultimately used as reference, which sets the reality bar for our work. The buildings that needed to be on fire were partially set alight so we had great reference for our CG fire. Likewise, the SFX and stunt teams did incredible work with the fires and explosions outside the gate – stunts really ran through those explosions and caught on fire. We used our proprietary Loki solver to match those fires and explosions – we got close enough that we could lay our fires over the top of the plate where we needed to for interaction purposes and expanding on what was shot. The shots where Tommy is shooting down into the fire line, for example, those are a mix of practical and CG. The reverse angle looking at Jackson, the explosions, and fires there are all CG. Special shout to Tomo, Duarte and the FX team for their relentless effort on the fires and explosions. We had a really solid tool for defining which Infected should catch on fire, what sort of coverage they should have, detecting what sort of motion we were dealing with, and appropriately driving each fire simulation based on that – in some cases we had almost 400 Infected on fire!
The snowy environment surrounding Jackson was a mix of full CG and extensions. How did you ensure seamless integration between the live-action plates and the digital environments, especially with all the added vegetation and snowfall?
NE // We generally replaced the ground from right outside the gate onwards. Even shots where we were close up on the feet of live action Infected, the artificial snowpack of the set wasn’t as deep as intended so we added CG divots and kick up to those. Distillery had done a great job with Jackson in episode one, we were lucky to inherit their buildings and the mountains behind the town, but looking out from the Jackson gates through the valley (which also happens to be the name of the episode) – that was all CG, some great work led by comp supe Mark Julien and environment artist Bjoern Kaehler.
Can you tell us more about the chase sequence with Abby? What were the most complex aspects of building and animating the environment for that scene?
NE // This was an interesting environment that had different requirements throughout. When Abby first rolls down into the basin and realizes there are a bunch of frozen Infected there – we added bodies, snow mounds and additional tree cover. When the chase kicked off, it got a little more complicated and we ended up with some pretty hero tree additions, as well as the copper mine and mountainside, which were all CG, as were the high wide shots. They ended up looking great thanks to the reference plates Alex shot in Kananaskis, as well as the talent of both our environment lead Heeeun Kim and concept artist Karl Sisson. The snowpack was also at its deepest in this environment, so as well as our CG horde there was a lot of snow interaction in every shot.
Can you tell us more about your Nuke-based system for controlling weather using machine learning and FX simulations. How did this system change your compositing workflow, and what advantages did it offer on a production like The Last of Us?
NE // I think for any big show where you are going to hit a point of critical mass in terms of shot production, you need to have processes in place before that point so you’re not having to solve the same problem over and over in every shot. It was clear both from the storm really being a character in the episode and from the wildly varying shooting conditions on set, that conforming the weather to the intended blizzard conditions needed to be a robust process. The machine learning part of it enabled depth extractions from monocular plates and also gave us base segmentation, both critical for inserting our CG horde and weather into any plate. The depth extractions were remapped based on deep data from set lidar and CG characters, giving us a cohesive and stable depth space to work with. On top of this comp supe Tobias Wiesner and lead Owen Longstaff developed a number of Nuke-based tools for blowing snow, volumetrics, embers and ash, and even gore hits and spatter. Compositing artists would identify a wind direction based on the plate and the system would populate a shot with a really good starting point for the weather, with full artist control for tweaks on top.
Your pipeline also featured dynamic clothing and hair refitting, along with procedural texturing. How did these tools improve your ability to iterate and respond to client feedback, especially across such a large and detailed crowd?
NE // Variation was another big problem for us to solve. We had almost 30 base Infected we built from scans of the stunt actors wearing prosthetics. Many shots required over 100 CG Infected seen at almost hero level, so we knew we would run out of variation with just our base builds. We came up with a mix and match system that would allow any Infected to wear any piece of clothing and hair. The dynamic refitting of those 400 or so collections was something our Creatures lead Claudio Gonzales figured out and it’s hard to overstate how important this was – as our Infected were based on varyingly proportioned scans, we were looking at manually refitting all those collections without this system. It operated per character in the pre-roll of our simulations to balloon each garment and then shrink-wrap back down to conform to the intended proportions. Importantly too, the 400 collections we identified all had proxies, so Layout artists could generate quick playblasts to judge variation and easily swap wardrobes where repetition was noticeable, all without needing to render anything. The procedural patterning tool was built by our Lookdev supe Emiliano Padovani and CG supe Snowy Leo – this allowed lighting artists to have full control over the clothing shading for any Infected, and when we started getting notes like ‘can we have someone wearing a yellow checked shirt over there, and maybe red stripes over here’ it was great to be able to confidently say ‘yep no problem!’
Looking back on the project, what aspects of the visual effects are you most proud of?
NE // The dogs in episode two turned out great, and ideally you can’t tell how many infected were added to the subway in episode four or which Stalkers were CG in episode five. That said, there were a few siege shots that initially felt borderline impossible. One was the moment outside the Jackson gate when the horde suddenly changes direction and runs over mounds of bodies. Another was the ‘conveyor belt of death’—as they round the corner onto Main Street and just keep coming while being mown down. Those shots were the culmination of everything we anticipated and prepared for earlier in the process, but they were still daunting to execute.
Honestly, what I’m most proud of is the team spirit, artistry, and perseverance that everyone brought to the project every day. That’s what made shots like those possible.
DY // It’s hard to pinpoint one thing as there are so many things to be proud of…. The Bloater moments turned out excellent, and many aspects of the Infected crowd acting as an organic horde was also very difficult and rewarding. I guess what I’m most proud of is the invisible work we did. There are several moments in E02 that integrate so well people will be surprised that the Infected were CG.
How long have you worked on this show?
NE // Around 15 months.
DY // I started in the beginning of 2023 and ended in early 2024.
What’s the VFX shots count?
NE // I think our final count was 384.
A big thanks for your time.
WANT TO KNOW MORE?
Weta FX: Dedicated page about The Last of Us – Season 2 on Weta FX website.
Alex Wang & Fiona Campbell Westgate: Here’s my interview of Alex Wang (Production VFX Supervisor) & Fiona Campbell Westgate (Production VFX Producer).
© Vincent Frei – The Art of VFX – 2025
