The Lost Bus is a 2025 survival-drama directed by Paul Greengrass for Blumhouse Productions in association with Comet Pictures and Apple Original Films. The film is based on the non-fiction book Paradise: One Town’s Struggle to Survive an American Wildfire by journalist Lizzie Johnson. Set against the 2018 Camp Fire in Paradise, California, it follows a school-bus driver and a teacher who fight to guide 22 children to safety through an encroaching inferno. Combining Greengrass’s documentary-style direction with large-scale visual effects and environmental reconstruction by RISE FX, the film depicts one of the deadliest wildfires in recent history with stark realism.

Supervising the inferno: Oliver Schulz (IMDB | Linkedin) is a senior Visual Effects Supervisor at RISE FX, the Berlin-based VFX studio. Over more than a decade at RISE he has guided VFX supervision on major international productions including Fantastic Beasts: The Secrets of Dumbledore, Blue Beetle and Megalopolis, among many others.

His background spans concept art, digital matte painting and 3D environments; skills that helped him to take the creative lead on blockbuster-scale environment and FX heavy shows. In this interview he reveals how he and his team at RISE tackled the challenge of recreating a burning landscape for The Lost Bus, combining procedural geography, wind-driven vegetation, deep-rendered volumetrics, and machine-learning techniques for depth integration, to bring the inferno to life on screen.

DP: How did you get onto the Lost Bus? 

Oliver Schulz: I came on board after wrapping up on Megalopolis and jumped onto the very first meeting with Charlie Noble and Gavin Round, Production VFX Supervisor and Producer. The project was already awarded at that time so we directly started talking about the Sequences and the scope of the Rise portion of work.

Luckily or better sad tragically this was a real event so in regards of look, there were many references and documentation of this day. So our first meeting was looking through a lot of real world footage from all available sources. Charlie had been prepping reference reels from the very beginning, so we could hone in on a lot of specific ones for each portion of the work because he had references for all of them!

DP: Roughly how big was the RISE team on The Lost Bus, and how long did you spend from first build to final comp?

Oliver Schulz: We started with a very small core team in May 24 and delivered the last shots at the beginning of 25.  I think around 50-60 people worked on the show in total during the production with up- and downramping based on specific project needs like temp deliveries for example.

DP: The antagonist of the movie is the geography of a very particular area, and the fact that it is on fire – how did you make sure that it was recognizably that specific part of the world? 

Oliver Schulz: We started with real world data derived from elevation models. That gave us a pretty good grounding in reality. We got lidar scans for very specific locations like the Pulga Bridges for example which was invaluable as this is usually something one doesn’t get from any publicly available sources. We spent quite a bit of time to get us a very good foundation of all key locations, which meant that everything had a geometrical base until the very last mountain you see on the horizon.

Our Lidar Supervisor David Salamon was instrumental in setting up this base. He used some maps imagery to give a rough base color to all those individual geometries that served as a rough guide later on in layout and surfacing for distributions of materials or assets. One has to keep in mind that most data was post 2018 so for instance vegetation had to be recreated from mostly photographic references shot before the fire. We tried to stay as true as possible to real world geography, but later on of course things had to be changed for storytelling reasons. 

DP: When merging New Mexico plates into your California canyon builds, how did you maintain scale and geological continuity?

Oliver Schulz: As production did not have access or in case of some very sketchy roads didn’t shoot at the original location for safety reasons, they did some scouting through the US and chose some New Mexico locations as stand ins for some of our sequences. The most prominent was the Pulga road for sure. In the film, the first responding firefighters trying to get to the origin of the camp fire first get sight when they are on top of the Pulga Highway bridge crossing the Feather River Canyon.

Without any better option they decide their best shot is to try and get to the fire following a very narrow road on the slope of the canyon. All shots on Pulga Bridge were shot on a Studio Backlot featuring full CG Environments including the FG bridge. This narrow path however was all shot on the New Mexico location with two big fire engines driving a slightly wider road. 

In production that meant that we replaced most of the visible Environment due to a couple of reasons. First of course it needed to have the right roadwidth and the correct canyon in the background. Second we needed to have very windy vegetation everywhere. Third in case all of that worked in camera which was pretty rare we still needed to put FX Elements into every shot consisting of dust, debris, smoke and later also embers. With those guidelines in place probably 90% of the shots became full CG exteriors only keeping small bits of photography for fire engines and some road pieces.

Once all those went into layout we made sure to keep a senseful progression to those shots meaning having the firefighters travel along the road during those shots in cutorder. The topography of the shooting location though was pretty different from the storypoint progression on the pulga road, so was the framing in camera when pointing at the fire from the fire engine interiors.

That of course meant as good and real our base was, it needed to be heavily augmented to make sense with storytelling and framing choices. Most shots feature the correct BG canyon but the midground is totally made up to allow for good view to the fire origin. All of this had to be understandable even with very frenetically moving handheld cameras.

DP: Let’s talk vegetation: How much botanical creative freedom did you have? 

Oliver Schulz: Vegetation was a big part of the environment and one of the reasons for the rapid spread of the fire. All FVX vendors had to tackle it in one way or the other which meant all were contributing to the research for which plant goes where. The foundation was once again the research and material collection from Charlie and his team.

We focused on the most common species found in this region of CA and made sure the level of dryness and the distribution made sense. So in this sense there was not too much freedom here as everyone tried to make this as real as possible from this point of view. For the build we actually just used the most common ground which is Speedtree with some augmentations done in Houdini. Part of the assets were also shares from other vendors which just needed ingestion and rigging in FX.

DP: You mentioned building a hierarchical “ecosystem” in Houdini. How modular was this system, and how much hand-authoring did artists still need to do per shot?

Oliver Schulz: This was something we actually invested a bit of time in at the very beginning and was overseen by CG Supervisor David Schulz and Layout Lead Mareike Loges / Senior Layout Artist Björn Markgraf. The core idea is nothing new and hierarchical just means that you start from the biggest Elements in your kit and than go smaller and smaller based on the previous distribution of Elements. First step is to either scatter or handplace big trees for example, following this you end up with a certain distribution. 

Based on this the system places smaller entities like younger trees or seedlings and smaller shrubs and bushes around or between the big trees. This distribution is based on simple rules like distance or terrain steepness. In case of the Pulga road we divided it up in two categories: mountains and roads. Both had similar procedures. We would always start with the rough blocking geometries matching either scan data, elevation data or sometimes just made up. From there we would generate the base coverage of rock cliffs which would hold out trees mostly in those areas. Following this we created the trees and bigger vegetation which would determine the ground coverage of rocks vs more pepply ground.

Roads were pretty similar but less complex as they mostly feature small stones. Again here we used some manually created maps to drive the distribution of small vs bigger pebbles that mostly accumulate on the side for example.

The toolset itself worked pretty well and as it was applicable as a template we could have a fully laid out shot in a day. Shot specific adjustments were applied on almost every shot though, mostly for continuity, visibility or art direction purpose. 

DP: Vegetation, environment, and FX were all dependent on one another. How did you keep versioning sane between departments?

Oliver Schulz:  That was a big topic indeed and it only is possible with two things: a rigorous approval system and a good pipeline that helps you track those approved layouts. We rely on our usd pipeline to do exactly that for us, it makes it somewhat easy (Im sure layout and production will hit me for this) to track department versioning. For each layout update we`d always get automatic QC renders that run through our inhouse “slapstick” system which is our inhouse auto comp engine.

When Layout would do a specific change it would publish this either on a shot or on an Sequence/Environment level. This will trigger a QC render from the shotcam of the affected shots. Once the rendering is done it will have a postjob that combines it with the prepped plate in nuke and runs another renderjob that will give you the layout reviewable which is than checked and can be approved and pushed into the pipeline from RV. This Layout than becomes available to the FX department which would run all needed simulations and hand off another QC reviewable for approval. Without those systems in place it would have been a nightmare to stay on top of all these versions!

DP: You said the layout department drove wind direction and strength instead of FX. How did that change your creative workflow?

Oliver Schulz: Yes that was indeed true and one of our early conversations we had internally to determine the approach all Environment builds would share. It was a practical decision based on two factors: We would simulate all vegetation on the asset level in different windspeeds for efficiency reasons and we wanted to keep iteration loops to a minimum.

This meant that I wanted to look at layout versions with moving vegetation as the strong directional wind would make it necessary to consider this already while layouting trees. As the direction is clearly visible you cant rely on spinning a tree 360 deg free in Y to create variations as the direction is “baked” in, so you need to see it moving in order to determine if an environment looks good!

The second reason is that FX needed to take care of vegetation simulation only once and when approved never needed to come back. This system worked really well and was accessible to the layout artists down to the single blade of grass, meaning one could really art direct where and how much specific things should move.

DP: You divided smoke into “hero” and “residual” categories. How did you manage density and readability without losing visual clarity?

Oliver Schulz: These two categories were simply based on the fact that we needed to deal with smoke in almost every shot. The ever present residual smoke needed to inherit a direction, needed to be art directable and also needed to render as fast as possible. Its pretty much the equivalent of atmospheric perspective in a wildfire scenario. Our Fx Supervisor Akin Göcmenli came up with a system of instanced presimulated caches that sometimes could consist of thousands of individual ones.

We started by doing simulations of smoke with a constant wind direction and speed that had a hidden source of emission and diffused pretty soon. That gave us a very soft falloff to the edges of the simulation grid which made these simulations perfect to overlap and look as one big single instance of smoke. On top due to those aspects it was quite easy to remove single containers and punch holes into the wall of smoke for visibility. We also invested a bit of time to develop shaders and render efficiencies to cut down on notoriously long volumetric rendertimes for this element.

Hero smoke was the category which either had a visible emission source in frame or simply had a hero storytelling element. These were usually shot or sequence simulations as they were mostly much denser and most of the time also much closer to camera. We also spend a good amount of time matching shading and simulation to real world references.  The secret to readability also lies in relentless QCing of outputs to make sure once you kickoff the expensive lighting renders, you are as certain as possible all elements are going to work. 

DP: The ember work looks incredibly detailed. How did you simulate believable motion in strong winds without visible repetition or looping patterns?

Oliver Schulz: Embers were a big part of the equations so by now you might guess…. Yes we spend a bit of time in asset prep to build some solid foundations. The first thing is of course the driving factor for all fx aspect: the wind.

FX developed hero wind forces that we used to simulate all elements with. A good amount of chaos and variance is key to not run into issues with readable patterns in any simulation. Another factor is collision. Embers will behave a certain way when they collide and thats what we tried to replicate. Also the ground plays a big role especially with the heavier emberclumps that slide over it.

Reality is unbeaten when it comes to little quirks and anomalies especially for something as complex as this. As no one is able to have a ground as detailed as the real world we also sometimes had a collision geometry that had slightly more displacement in order to have more detailed collisions happening. 

DP: Lighting and volumetrics are natural enemies. How did you maintain physically plausible lighting through that much smoke and fire?

Oliver Schulz: That was one of the biggest questions going into this project, considering that what was shot on set sometimes had very little to do with what ended up on screen, especially in terms of atmospherics. The best base for something to look real is to match the real thing. We did so in our asset phase and made sure our shaders and lightrigs were physically plausible, especially the ones only used to develop assets. 

We would match greyballs and reference macbeth charts in order to make sure scene lighting was correct in terms of lightbalancing. Then from there we developed shaders for all aspects.One of the most common issues I see is that volume and surface renders aren’t lookdeved in conjunction. What you end up having to do is to grade surface and volume render differently. This leads to very unrealistic renders very fast because there is no ground truth you can come back to.

We tried to make sure all our shaders worked with each other to have exactly this common base. Also when dealing with dozens of light sources and those issues on top that’s definitely a position you don’t want to be in when time is running…. All volumetrics do have very different properties to them where one of the biggest is how they scatter light. Back to front scattering can take a volume from being ultrabright to consuming all lighting energy and being pitchblack. So once you matched the real thing, use those tools wisely to deviate from there and support the story. 

We tried to always start with a balancing pass usually done still in lighting. This goes to comp as the foundation to do all the finetuning with. Still there was a lot of tuning left for comp and also we needed to break reality more than once to make sure that what you wanted to read in a frame remained readable when tons of smoke and fire went in front. Sometimes we needed to go as far as use the deep data to pull things in and out of the smoke to make them visible. Still the most valuable tool you have is the artist’s eye to determine the sweet spot of good vs real.

DP: You mentioned deep rendering bottlenecks, like OIIO running out of patience with too many AoVs. What exactly went wrong first?

Oliver Schulz: Haha yes that was one of the issues that came when switching to full deep with our renders. That means you have each component of each lightgroup rendered with deep data to put it back together in comp. That resulted in a lot of channels that apparently were too much for OIIO to handle. Thank god that was fixed otherwise I wouldn’t write this story now ;)

DP: Rendering holdouts took up to an hour per frame. Did you develop any automation or optimisation to make deep rendering less painful?

Oliver Schulz: Indeed rendering times for deep holdouts were quite painful and nothing to speed up really. With so many volumetric elements you need to deep hold out everything with everything to make sure its accurate. If you multiply this with the number of separate elements rendered and with the amount of light Aovs times the amount of components you end up with a staggering number of renders.

Plus in the end you need to denoise all frames so the best solution was try to plan out delivery dates as good as possible to have time for all those thousands prerenders to run on the farm. Still our compositing Supervisor Oliver Hohn and Lead Nicolas Burgers had some longer evenings ensuring all renders were there the next morning to be picked up by the compositing Artists.

DP: You used machine-learning depth generators to create deep data from plates. What tools powered that, and how reliable were the results?

Oliver Schulz: We started testing tools quite early in anticipation of very challenging compositing work. DepthAnything v2 was what we ended up using as a default prerender pass. The results were a mixed bag considering the wide range of plates we worked with, although it proved to be valuable to have. Comp remapped the relative value output of the depth passes to absolute values from deep data with help of lidarscans or renders and was able to create some good integration especially with more wispy type of smoke.

For denser smoke and more accurate holdouts especially for actors we still needed to rely on a lot of manual roto for good integration. The AI passes proved to be pretty successful though for fast temp work as you get something going in no time. Issues were mostly the missing good temporal stability and also the lack of precision. 

 DP: Before deploying new tech like ML depth generators or procedural ecosystems, how do you test them safely inside production?

Oliver Schulz: We implemented those during production directly on our project infrastructure, so developed, tested and used simultaneously. 

DP: Were there any spectacular ML depth map failures, like smoke reading as solid or background cliffs collapsing?

Oliver Schulz: Depth popping or lost shapes were the most common ones. But as none of these passes were used without correction in comp I’m sure I haven’t seen all of them!

DP: With so many volumetric layers, how did compositors manage complexity without drowning in passes?

Oliver Schulz: We have standard workflows for loading CG renders into nuke which do provide a basic level of organization. However the more elements you have the bigger the compscripts and we had some good ones for sure!

DP: Fire colouration is tricky. Did you use any spectral rendering or rely purely on LUTs to match on-set lighting and heat distortion?

Oliver Schulz: That is very true.. Luckily production tried to shoot everything with a practical fire which provided a good level of references in camera. If you than try to render as physically plausible as possible and have something in frame that you can match exposure to you are already halfway there. We didn’t use any spectral rendering here and rendered everything through Houdinis Karma in RGB.

DP: You switched to full motion-blur sampling for embers instead of faked streaks. How much did that impact render time, and was it worth it?

Oliver Schulz: Oh that was worth every minute of rendertime.. Real Motionblur for an element which is mainly visible in motionblur is a good investment. Plus the rendertimes weren’t actually that bad and took only a couple of minutes as you are not dealing with an expensive shading as well. The biggest benefit is getting nice curved and very interesting blurs especially with collisions.  The trick actually is to only invest time where its needed and render other elements with less costly settings. Deep compositing allows for it as you are not bound to any holdouts and you can combine differently rendered motion blur without any problems.

DP: How crucial was RiseFlow for distributing simulations and maintaining consistency across all sequences?

Oliver Schulz: We started implementing RiseFlow at the very beginning once we had our initial workflow for distributing elements figured out. The development was done by our Head of Pipeline Paul Schweizer and the implementation on the show was spearheaded by Jonas Sorgenfrei.

It actually is a very versatile framework that we use for a variety of tasks here at Rise. Its a modular node based System that can take arbitrary inputs and execute them in a chained workflow. FX built templates for various scenarios that got exposed variables like wind direction, speed, inputs for collision geometry etc. These could than be varied per shot and sent to the farm for execution. Once all those Sims were done, QC renders were submitted to the Farm and when completed, auto comped in Slapstick.

That meant that one artist could do changes on a big number of shots by adjusting the template and than resimming and rendering them over night. All render elements were deepcomped with our deep plate workflow and reviewed the next morning. This allowed for rapid adjustments and turnarounds which was a very crucial aspect of this fast paced production. 

DP: How did RiseFlow and Slapstick communicate between departments for reviews and dailies?

Oliver Schulz: RiseFlow and Slapstick are two different things really. The point where they communicate is that Riseflow might trigger a farmjob where Slapstick is hooked in as a post process that gets triggered after completion of the render. Slapstick again is a modular node based system implemented in Nuke that allows for a generalized template to be created. These inputs could take for instance all general elements that comp might use to layer a shot like mainplates, rotos, colorcorrections, lensdistortions etc and comp them together. We use Slapstick in all departments to create automatic reviewables for assetbuild like turntables with reference images, lighting slaps,fx slaps and so on.

DP: You’ve called The Lost Bus the toughest matchmove job you’ve ever seen. How did you solve the handheld, wet, low-light camera challenge?

Oliver Schulz: That was a tough one indeed. To solve this it really just comes down to the excellence of all individual artists that created those matchmoves. So there is no magic recipe to get through so many challenging matchmoves…

DP: Greengrass loves long, continuous takes. How did you manage to iterate and render efficiently on such heavy, unbroken shots?

Oliver Schulz: I guess it is really to choose your battles wisely… Invest into a good foundation early on and make sure to be as precise as possible in prep phase. Once the show is running and you are in full delivery mode there is no time to go back and redevelop anything.

Render optimization as much as possible and then relying on everything that was set up in the beginning is key to not have to think about accuracy anymore when you are trying to finish the shots. We did this and it really paid off, though having a couple of long shots with lots of elements to render we never ran into the issue of having to fear a render didn’t get finished in time.

There were some challenging shots for all departments involved but again the prep phase paid off and we managed to deliver everything in time. It’s really a Situation in which the Production team led by Michelle Cullen and Production Manager Androniki Nikolaou outdid themselves by planning and scheduling every milestone in production to make sure we had what we needed to finish shots in time. Of course that also means adjusting and revising this schedule each and every day based on client comments and changes.. It’s a tough job to make sure the whole production runs like a well oiled machine!

DP: Deep compositing only works if all layers align perfectly in space. Did you use diagnostic tools or pure visual QC to verify deep accuracy?

Oliver Schulz: The good thing about deep is that it’s pretty accurate as long as the sampling increments in depth are small enough for certain elements. It’s a game of keeping error thresholds low enough so you don’t pick them up actually. The balancing is precision versus filesize. Surface renders aren’t an issue really as you are dealing with front and backsides of hard surface objects really.

The fun starts with volumetric elements and this is where you need to tweak the settings a bit to make sure you don’t end up having 5GB per frame in volumetric renders. Still frames could grow to well over 1GB on bigger shots with all elements included, so we needed to do some rough calculations beforehand to make sure we weren’t running out of allocated serverspace.

DP: How did you maintain consistency for fire behaviour across sequences? Was there a single reference look, or did it evolve shot by shot?

Oliver Schulz: One of the big topics obviously here as the quality of the fire not only needed to remain consistent but also serve the story in how it behaves. When you look at fires in reality they all have very different qualities to them depending on a ton of external factors like what is burning, where it burns, what is the actual heat it produces and what is the influence of the wind and so on. So yes it’s crucial to pick a reference and not try to incorporate them all. The initial tactic we used was to create asset based fires with all components.

The two types that production defined as assets were “spotfires” and “forestfires”. Pretty generalized in description though mostly divided up by scale. So we took those two types into asset development and created a little scene with them. Forestfire in the background and spotfires in the foreground. This scene actually was the same one we used to lookdev all assets in. So we had a common ground for all assets really and the fx ones were not different.

We picked a general reference we felt was working well for each category and supplemented that with references that production had shot on set. The shot element though were mostly run by gas so wouldnt really emit any smoke but were a general ref in terms of breakup and edge qualities. Also those would come in native resolution where most of the actual refs are cellphone captures of much poorer quality.

So with all those references in place we started matching the fires again in different windspeeds. We tried to also implement all little details especially on the bigger forest fire like flambursts on dry wood, falling burning pieces of wood etc. Once fire was in place we hooked it up with all secondary elements like smoke and embers. We had already pretty robust setups developed for each of them individually so we could already build on a solid foundation using those as a base.

Once this little scene was successfully approved by production to go into shots we splitted out the individual components as assets again. These had all elements attached like smoke emission, ember emission, lots of different masks for heat distortion and were ready to be dropped into shots.

Using this technique we had a very solid foundation of very similar looking and behaving fires. Of course for hero shots we would need to resim those, but with setups in place and our template system it was mostly straightforward. Of course there are shots that need to tell a certain story like a fire coming right at you towards camera. Solving a problem like reading a perspective of a selfilluminating matter coming right towards camera is a different beast though you can’t prep for! This just takes a lot of creativity and trial and error to get right…

DP: What’s the single biggest creative takeaway from The Lost Bus you’d carry into your next show?

Oliver Schulz:  Don’t try to put out all the fires at once….

DP: Which shot makes you proudest or gives you flashbacks?

Oliver Schulz: Oh there are so many good ones really, honestly when I was watching all shots in a row I was so happy about the overall level of quality the team achieved in every aspect. So hard to pick singles but the Embercam full CG shots looked amazing on the big screen and were pretty spectacular… but getting them to the state we delivered them in was quite a journey…

DP: Finally, if you had to redo The Lost Bus from scratch, what would you rebuild first? Vegetation tools, compositing templates, or your caffeine reserves?

Oliver Schulz: Myself :)

The post Rendering the Inferno at RiseFX: The Lost Bus first appeared on DIGITAL PRODUCTION and was written by Bela Beier.

A tasting of the dynamics and geometrys side (of) effects

This release is not about the number of features, but about finishing what was started, achieving true production readiness, robustness, performance, and ease of use. It’s a version focused on quality of life. Feature sets like MPM simulations and Karma have matured like a good wine. The machine learning tools respect the artist’s skill set and have left their state of play are actual usefull in production.

Even though SideFX remains modest in its official statements, Houdini 21 is a massive release. For the feature-hungry among us, the highlights include a fully matured APEX; a refined and clever animation and rigging framework; a simulating Copernicus (think Substance on Dope), and in the VFX realm, the production-ready MPM Solver. On top of that, we’re seeing machine learning tools popping up in all the right places (AI for people who don’t expect a Pixar film from a single click), faster and expanded rendering with Karma, and a healthy dose of quality-of-life improvements.

Since our editorial cat can only count to 300, we can’t tell you the exact number of new or improved features, but we’re impressed nonetheless. And soon, you will be too. To avoid overwhelming both writer and reader, we’re breaking this article into a mini-series. We’ll start with VFX & Geometry, followed by Solaris & Karma, Copernicus & Terrain, and the massive Animation & Rigging tools.

MPM

One of the most exciting additions in H20.5 was without question the MPM Solver. The MaterialPointMethod truly allows you to simulate a wide range of different materials. From water, snow, and sand to honey, metal, and concrete, all within a single Solver setup. The geometry is simply substituted by points or particles, which are then simulated. The different materials interact physically accurate and constraint-free, purely based on their assigned parameters.

The initial release was already impressive, but it left users standing somewhat in the simulated rain when asking: “How do I get a proper mesh with UVs?” That exact issue has now been beautifully solved in H21. Any MPM simulation, whether rigid body or fluid, can now be meshed (as polygons or SDF volumes), including UVs, color, and other attributes, using no more than two nodes.

A nice and usefull addition to this workflow is the MPM Debris node, which generates new points along fracture lines as sources for smoke, dirt, or secondary debris effects. So let’s take a look at meshing hard and fluid or granular surfaces across a few setups and scenarios and wrap things up with a creamed cookies drink while watching the official and very excellent MPM demo.

Surfacing MPM Simulations

Testing time! For this test setup, we’ll have a clay ball smash into a vase model, paying close attention to UV transfer and the generation of smaller fragments through the Debris Source. The easiest way to start an MPM simulation is by typing “MPM Configure” into the node search. This gives you a complete set of starter nodes right away. (Under MPM Configure, you’ll also find plenty of additional example setups for study or creative repurposing.) By the way, the MPM container on the far right controls the overall resolution of the entire simulation.

We replace the default sphere with our own model and can now assign materials directly inside the mpmSource node and tweak them to our liking. It’s genuinely fun. Feels a bit like a mini-game. Since a concrete material would be more realistic but also quite boring as a vase material, we went with Chunky Snow instead. The environment comes in via an Object Merge directly into the mpmCollider Sop, ready to go. Our clay ball, the antagonist of this little simulation, isn’t a collider per definition but another mpmSource ready to be smashed, merged with the vase and given its own material behavior, Chunk Soil.

To make sure the particles can actually see and “love” each other, we need to enable Particle-Level Collision in the solver. The new Auto Sleep feature helps keep the vase passive until the collision happens, preventing it from collapsing at frame one and saving quite a bit of compute time.

Since the clay ball will be meshed as a fluid and granular surface and the vase as a rigid object, we first separate the MPM particles using a Blast node, filtering them by their respective source names.

For surfacing the vase, we use the mpmPostFracture node, which takes the rest-geometry and the MPM particles as input. This node essentially breaks the geometry apart “end to start,” so it needs to be fed the final frame of the simulation. After that, we choose either Voronoi or Boolean Cut as the fracture method. The latter can generate interior details, subtle irregularities on the inner faces of the fracture, that weren’t visible before, and it’s also faster to compute. We can further control the number of pieces as well as define the minimum fragment size at which new pieces should be generated.

The final node in the chain is mpmDeformPieces, which transfers the newly generated fracture geometry onto the MPM particles and just like that, the vase shards, physically convinced they’re made of snow, go flying through the scene, complete with perfectly intact UVs. For more of a muddy mess, we could have generated a liquid or granular surface instead, but we’ll save that for the clay ball. The result from the Debris Source, which lets you precisely define when and where particles based on fracture are emitted, is then passed into a POP network, including collisions from the vase and background).

Time to get serious:

Continuos Emissions & Surface Tension

With this new option, you can quickly fill containers, simulate expanding materials, or layer different materials on top of each other. The option is located in the mpmource node and spreads new particles apart using positive pressure.

Let’s take our vase and let a thick, viscous something ooze out of it. A good chance to show how simple MPM can be: if we want things to float inside the fluid, they just need a lower density. A few cubes are generated and assigned a jelly material. Two geometries, two MPM Sources, merged and fed into the solver. That’s all it takes. Just as easily, you can mix different fluids within the same setup. And for a bit more drama, we can dive into the solver and add a POP Wind node with some turbulence.

Surface tension allows for realistic effects such as droplets, tendrils, and flowing water. H21 introduces two new ways to control surface tension for both liquid and viscous materials inside the MPM Solver. The Point Based method offers higher accuracy and stability, making it ideal for small and detailed simulations. The Grid Based method, on the other hand, is optimized for performance and handles millions of particles more efficiently, which makes it better suited for large-scale scenes. External forces and friction can be increased if objects keep moving when they shouldn’t. Otherwise, you might end up with a scene straight out of Terminator 2.

To achieve high levels of detail without an unnecessarily large number of simulation points, the new, more precise collision detection allows you to use the fracture edges of a simulation as the source for a targeted secondary simulation. The attribute Jp (Plastic Compression/Stretching) is key here. It can be used to isolate the fractured areas and feed them into a Surfacing node set to VDB mode. This resulting volume can then serve as the source for the second simulation. And don’t forget to use the main simulation result as a collider.

And finally to top it off, the official demo. A detailed breakdown of that beautiful cookie shot.

Machine Learning in Dynamics

You won’t find any generative AI native in Houdini, but rather a growing collection of smart, locally running models, often trained by yourself, designed to simplify or speed up time-consuming tasks.

Surfacing Flip, Vellum & Particle Simulations

Alongside Neural Point Surfacing for MPM, the new Neural Point Surface and proven Particle Fluid Surface nodes now bring neural meshing to FLIP, Vellum, and POP simulations as well. Until now a bunch of points are trying to reconstruct the surface of a material. With neural meshing, you can now achieve much sharper, more detailed surfaces across a both high and low frequencies. The result: surfaces that are no longer uniformly fuzzy, but crisp, structured, and temporally stable. As before, you can train your own models, but even the included presets already produce finer details. And thankfully, the whole thing is GPU-accelerated.

Volume Upres

The core problem behind volume up-resing: for efficiency, artists often create and approve low-resolution simulations. But once the voxeldensity is increased, the overall shape of the sim tends to change. With the new tools, a low-res simulation can now be upscaled while preserving its exact shape. This not only keeps previously approved versions intact, but also allows for far more iterations. A model that’s been trained on a specific motion or behavior can now upscale multiple variations of it.

The Billowy Smoke recipe (or shelf tool) comes with a pretrained up-res model already integrated. Let’s start by looking at a comparison between the low-res input simulation and the 3× up-res result.

Details are nicely added, while the overall shape is preserved impressively well. Caching took a bit of time, but it’s still far faster than running a full high-res simulation. The real idea, however, is that we can now reuse this up-res model for all our future Billowy Smoke setups and honestly, we probably should. So, let’s quickly modify the setup and see if we can break the upscaler.

This time, the solver runs at a voxel size of 0.05, using the 2× upscaler. The 3× model didn’t really add more detail, just extra waiting time. For a bit more fun, a Gas Wind with random direction and a collision shape were added. That collision shape, as it turns out, gives the upscaler a bit of trouble as seen below.

But when comparing the up-res result to a true high-res sim, it’s clear that the system is really good at preserving the base form. Doubling the voxel size in the high-res sim, on the other hand, changes the overall shape and eats up a ton of time but stays artifact-free. Or, if you really want perfection, you could just train a model specifically for this type of collision.

If you want to train your own model, this Equinox Hive talk walks you through every detail:

Zibra AI VDB compression

With this plugin you can save up to 99 percent of storage space when caching VDB simulations, which makes it perfectly suited for use in real-time engines such as Unreal Engine 5. The Zibra toolset, distributed via SideFX Labs, provides three dedicated nodes.

The first, zibravdb_compress, writes and exports .zibravdb files for use in Unreal and similar environments. The second, zibravdb_decompress, brings those files back into Houdini. And finally, zibravdb_filecache acts as a modified File Cache node that automatically handles compression, loading, and decompression for further use inside Houdini. Before diving in, you’ll need to download the model and obtain a license, potentially a free personal one if your revenue is below 100 K USD. The license management can be accessed directly from any of the three nodes.

For a quick benchmark, I used the Fireball Recipe and cached one regular VDB sequence alongside two Zibra versions with different quality settings. The original VDB sequence weighed 294 MB. The Zibra compression at a quality setting of 0.2 came in at only 5 MB, roughly 98 percent smaller. At a quality setting of 0.9, the result visually matched the original VDB almost perfectly while staying at just 36 MB, around 88 percent less.
That insane low file size at 0.2 naturally comes at the cost of lost detail, as visible in the comparison graph below. Still, the results are impressive and they open up the possibility for bringing volumetric simulations into real-time pipelines far more efficiently than before.

Pyro Shelve Tool Presets

Another way to get artists up to speed faster are production-ready presets, not just educational examples, but tools meant to be customized for your very real projects. Each of them comes with a ready-to-use Solaris network, fully set up for rendering straight out of the box. Even better, the Help section now includes a short guide and exploanation of important Nodes for every preset.

SideFX strikes a noticeably new tone here, aiming to flatten the learning curve rather than overwhelming newcomers with endless options (which, to be fair, they still do from time to time). These guides can be found under Documentation → What’s New → Pyro. In this section, we’ll take a closer look at three fundamentally different presets, each showcasing its own approach and creative use case.

Stylized Flame

With the refined Copernicus toolset, entirely new worlds open up: stylized fire based on a classic Pyro simulation, right inside Houdini. And if needed, even live-rendered in Solaris.

To put the claim of “easy adaptability” to the test, we took the Pyro Fireball preset and gave it a Toon-style makeover. Adding the VDB field “Flame” inside the solver’s output was all it took to make it work. The output from the Cop node, by the way, can be merged directly into the scene for further Houdini style editing.

Ground Explosion B

This shelf tool sets up a sparse pyro simulation featuring a large-scale explosion, smoke trails, and a shockwave. For more control and efficiency, it’s actually made up of two separate simulations, layered on top of each other and interacting through their velocity fields.

Candle Flame

A candle flame might not be the most exciting thing visually, but it’s one of those Pyro results you end up needing again and again. What makes this preset interesting, though, is the procedurally modeled candle that comes with it. Exploring that setup is almost more fun than the Pyro sim itself.

Thruster FX

In true H21 fashion and in the spirit of overall efficiency boosts, the new Thruster FX tool makes its debut: a setup designed to create engine and propulsion emissions with ease. It’s not just a new node, but rather a complete Recipe, a preconfigured network of nodes that some might, in hushed tones, simply call “presets.”

With a cheerful click on Thruster Exhaust in the Pyro Shelf Tools (or via Configure Thruster in the Tab menu), you’ll get a fully adaptable node tree, including a ready-to-render Solaris network. The effect itself isn’t a simulation but a cleverly layered, art-directable procedural system built around VOP Nodes. Multiple pyrothrusterexhaust nodes are stacked in layers, each responsible for different components like sparks, fire, and plasma. All working together to form a surprisingly easy to use thruster system.

So, what does it actually look like? And can it really be used straight out of the box, as promised? The short answer: pretty much, yes. The rendering comes surprisingly close to the viewport preview. To get it running, only a few connections inside the included Solaris network needed to be adjusted. For our small test scene, we did a bit of kitbashing inside Solaris, then added some glow and polish in Fusion.

Let’s take a closer look at the node, both from the outside and under the hood. The node expects a primitive as input, and a simple circle usually does the trick. It outputs both particles and a volume containing density and temperature fields. In the General tab, you can control speed, length, and the overall shape via a spline ramp. The Exhaust section handles the color ramp and lets you tweak the underlying noise pattern, which has a strong impact on the overall form. Under the hood, the node generates a VDB from Polygon, then modifies the result with a Volume VOP and a Volume Adjust Fog node.

As part of the ongoing effort to simplify things and lower the learning curve, SideFX has also released a good and detailed tutorial mini series: sidefx.com/tutorials/how-to-create-thruster-fx

Car Destruction FX

So that we can also crash the rigs built with the Car Rig SOP introduced in H20.5, Houdini 21 brings us the new Car Destruction Tools, led by the mighty RBD Car Fracture SOP, supported by the RBD Car Transform SOP. The first one takes care of fracturing and constraint creation, automatically handling the typical materials you’d expect in a vehicle: glass, metal, wood, and rubber. The RBD Car Transform SOP, similar to Transform Pieces, ensures that all pre-fractured parts are efficiently transformed based on the simulation points. You’re not limited to cars, by the way. Anything that follows the same basic logic can be blown apart. From motorcycles to helicopters, it all breaks just fine.

Destruction-hungry artists will find a detailed yet easy-to-follow example scene in the SideFX Content Library, the same visuals you might recognize from the keynote: sidefx.com/carbd-dual-car-collision/. A fitting go-through video can be found here:

Geometry, Viewport and other tasty QoL

Coming from its deep VFX roots, Houdini has taken quite a journey to establish its own distinct style of procedural modeling. With H21, that journey continues, extending existing nodes and adding a few genuinely useful new ones along the way. This time, even the viewport got some well-deserved love, now powered by Vulkan and capable of loading Gaussian Splats directly.

Sculpting in Time

The Sculpt SOP, introduced in H20.5 and (surprisingly) quite useful, now gets a genuinely groundbreaking new feature called Shot Sculpting allowing time-based, keyframe-free sculpting. Originally intended as a correction tool for character animation, the node turns out to be just as handy for VFX and motion design work.

Temporal control is handled through the Shot Sculpt panel, which at first glance looks a lot like an NLE timeline and, in principle, works much the same way. Sculpting can be organized into layers that can be offset in time, faded in and out (complete with easing), muted, or adjusted in opacity. Alternatively, you can use mask_track to paint time-based attributes, which can then be passed downstream and used in other nodes, for an obvious example, as masks.

Otherwise, the same rules apply as for the regular Sculpt SOP, whose updates we’ll take a look at next. In line with the new Shot Sculpting feature, the mask system has been reworked. Masks can still be painted manually, but can now also be loaded from an upstream float attribute, saved permanently, and blurred or sharpened as needed.

There are also new brushes. My personal highlight, the Elastic Grab brush:

Of course, the complexity and depth of ZBrush remain unmatched, but for many tasks, artists can now comfortably stay right inside Houdini.

Geometry Masks

There are also updates when it comes to masking. Several well-known nodes now include a Mask parameter, allowing the effect to be restricted to a painted or procedurally generated mask. Among them: Peak SOP, Soft Peak SOP, Inflate SOP, Flatten SOP, and Point Jitter SOP.

UV Flatten from Points

The latest addition to Houdini’s already powerful UV toolset could just as well be called “UV from Voronoi” since that’s exactly what it’s based on. The node distributes random or precisely placed points across the surface and uses them to calculate clean, non-overlapping UVs. It’s primarily designed for complex, high-resolution meshes, where traditional unwrapping tends to get messy fast.

Vulkan Viewport

Now enabled by default, the new Vulkan 3D viewport offers noticeably improved lighting, Ambient Occlusion, shading and ray tracing with built-in denoising, and a more accurate texture display though performance can take a hit if you push it too far. New worklights including a fully adjustable Dome Light, Physical Sky, and Three-Point Light setup now serve as the default viewport lighting.

Looking toward the future, the viewport can now display Gaussian Splats directly. Since splats are essentially just point clouds, and Houdini is fundamentally point-based, this opens up a rather promising combination. The .ply file can simply be loaded via a File SOP and passed into a Bake Splat SOP for further processing. From there, you can treat and manipulate the splats just like any other geometry using the usual SOP tools. More on rendering those Splats in the upcoming section on Solaris & Karma.

Curve Tools

The new Extract Contours SOP can generate object outlines from a camera’s perspective either directly as edges or as an edge group. Quite handy for toon-style effects.

The well-known Curve SOP now allows you to interactively split points into branches (with unique vertex numbers) or fuse them back together.

Unsubdivide

If things ever get a bit too much, this node can reconstruct a low-res input geometry based on Catmull-Clark.

Conclusion after some bottles of VFX

Even though the stated (and achieved) goal of H21 was mainly polishing existing systems and adding plenty of quality-of-life improvements, it still manages to sneak in a massive load of new features along the way. And we’ve only scratched the surface here. Deep dives on Copernicus, Solaris, Karma, Rigging, and Animation are already in the works.

What’s also refreshing is the ongoing effort to flatten the learning curve through better documentation, tons of in-house tutorials, and solid example files in the Content Library. Many things have become easier or let’s say, more accessible, without losing depth, at least for those who want to go there. As always, most nodes can still be cracked open and modified at their core. Nice one, SideFX.

The post Houdini 21: Like good wine (Part1, VFX & Geo) first appeared on DIGITAL PRODUCTION and was written by Manuel Kotulla.

If you’ve ever watched your pyro sim clip through a collider and thought, “Physics has left the chat,” this update is for you.

Headline Features

The big one is collisions: animated colliders and narrow geometry now hold up better, with smoke and fire retaining density and velocity instead of evaporating into numerical limbo. On top of that, the sourcing workflow has been redesigned, so setting up emitters feels less like solving a puzzle box and more like… well, just setting up emitters. The interface has also gone through a detox. Parameters are now sorted into sensible tabs, with advanced options politely tucked away, which means less scrolling and more simming. Performance has had a small but noticeable bump too, with simulations running between 5 and 15 percent faster depending on scene complexity. On the CPU side, fallback multithreading now relies on Intel TBB, giving more stability when GPU memory runs dry. And yes, it’s time to check your Houdini version: Axiom 4 supports Houdini 20.0, 20.5, and 21.0, with 19.x officially stamped “legacy,” which is dev-speak for “upgrade already.”

What It Means for Artists

The gist is simple: smoke looks better when it interacts with geometry, simulations tick along faster, and you’ll spend less time digging for hidden parameters. For artists cranking out pyro in production, Axiom 4 is less about flashy new features and more about shaving off headaches per day.

Pricing

Nothing has changed here. The commercial license stays at $199, the indie license at $99, and if you’re running Houdini Apprentice you can keep setting digital buildings on fire for free.

Outlook

Axiom 4 isn’t about reinventing the solver wheel but about refinement. With sturdier collisions, smoother sourcing, and faster turnaround, this release makes it easier to trust that your smoke won’t ghost through a wall. For production artists, that’s often the only feature that really matters.

Release Notes

The post Axiom 4 for Houdini: Faster GPU Pyro With Better Collisions first appeared on DIGITAL PRODUCTION and was written by Bela Beier.

Theory Accelerated has released a first look at Axiom 4.0, its GPU-accelerated sparse volumetric fluid solver for Houdini. The preview highlights more accurate collision handling, a rebuilt sourcing system, and 5–15% faster simulations.

Cleaner Collisions

Axiom 4.0 improves fluid–geometry interactions, particularly with animated colliders. Simulations show better volume retention and fewer artefacts in tight or complex spaces.

Sourcing Overhaul

The sourcing workflow has been restructured, paired with an updated Output tab for finer control. The new design aims to simplify parameter handling and make feedback loops clearer.

Performance Gains

On the performance side, Axiom 4.0 adopts Intel’s Threading Building Blocks for CPU multi-threading. This brings an overall 5–15% speed increase, especially when running outside GPU acceleration.

Release and Compatibility

Axiom 4.0 is in closed beta with no release date announced. It is previewed as compatible with “the latest Houdini builds” and the upcoming Houdini 21. The current stable release, Axiom 3.3, runs on Windows, Linux, and macOS. Pricing remains unchanged at $199 (commercial node-locked), $99 (Indie), and free for Apprentice users.

Industry Use

Axiom has been adopted as a faster alternative to Houdini’s native Sparse Pyro solver. It has already been used in production by Riot Games, Valve, and Muse VFX. The solver supports CUDA, OpenCL, and Metal, covering a wide range of GPU hardware.

The post Axiom 4.0 Sneak Peek for Houdini first appeared on DIGITAL PRODUCTION and was written by Bela Beier.

Freeing the VDB Pipes

Zibra AI has made ZibraVDB, its AI-powered volumetric compression tool for Houdini and Unreal Engine, available at no cost to indie artists and small studios earning or funded under $100,000 per year. This applies to both Houdini and UE5 editions, including for commercial use.

ZibraVDB compresses OpenVDB files, used for effects like smoke and fire, down to approximately 3 percent of their original size and claims to render them about twice as fast as Unreal Engine 5.3’s native Sparse Volume Texture (SVT) system.

What Changed—and What Remained

Until now, the free trial edition restricted users to compressing only five VDB sequences. With this update, the restriction is gone: indie users can compress an unlimited number of sequences and play them back in real time within Unreal Engine. Previously, individual subscriptions cost $10 per month or $150 per year. Those options remain for users outside the indie threshold, while studio-level subscription pricing is available on enquiry.

System Support and Availability

• Unreal Engine edition: Compatible with UE 5.3 or higher on Windows only. It’s distributed via Epic’s Fab online marketplace under a Standard license.
• Houdini edition: Compatible with Houdini 20.5.619 or newer, on Windows, Linux, and macOS. Available via SideFX Labs, an experimental plugin repository for Houdini users.

Installing either requires registering for a license key on Zibra AI’s website.

Final Note for Professionals

All claims of compression ratios and render-speed improvements are vendor-reported. Users should test ZibraVDB in their own environments before adopting it in production.

The post Zibra AI Frees ZibraVDB for Indie Artists (Briefly, Because Less Is More) first appeared on DIGITAL PRODUCTION and was written by Bela Beier.