Former Pixar FX Artist Reveals a Smarter Way to Refine Destruction

Destruction simulations can become extremely expensive to revise once their overall movement has been approved. Changing the fracture pattern or behavior of one area may alter piece indexing, constraints, collisions, and the motion of surrounding geometry, forcing artists to recalculate work that a director already liked. Former Pixar FX Technical Director Jae Jun Yi developed a layered Houdini workflow intended to make those increasingly specific production notes easier to address.

Presented at SIGGRAPH 2026, Multi-Layered RBD (Rigid Body Dynamics) Simulation begins with a relatively simple base simulation focused on composition and large-scale motion. Artists can then select individual pieces manually or through conditions such as size, velocity, impact strength, and collision events. Those pieces are fractured again, inherit the approved movement of their parents, and transition into a new dynamic simulation only when an artist-defined trigger is reached.

In this interview, Yi explains how identifiers, transform attributes, activation states, constraints, and collision adjustments maintain physical continuity between layers. He also discusses the workflow’s art-direction controls, its limitations for highly interconnected or interactive destruction, and his preference for extending Houdini’s familiar RBD toolset rather than creating a specialized system that other artists would struggle to understand.

First, can you please tell me a bit about you, your research, and your background as an introduction?

Jae Jun Yi: I have worked as an FX Technical Director for about 12 years, building my career across studios including Pixar Animation Studios, The Mill, and Blur Studio.

As an FX artist, I have always been interested not only in the visual result itself, but also in the technical structure behind it and in developing more efficient production workflows. In particular, I have been interested in how complex simulations can be controlled more efficiently while still remaining flexible enough to respond to a director’s art direction.

Based on this experience, I’ve also had opportunities in recent years to share some of my work and research at SIGGRAPH. In 2025, I presented a SIGGRAPH Talk on the sand simulation work developed for Pixar’s Elio. In 2026, I presented my research, “Multi-Layered RBD Simulation,” as a SIGGRAPH Poster, which is the work I’ll be introducing in this interview.

After leaving Pixar in 2025, I began working as an independent filmmaker, exploring the use of AI in animation production. In 2026, I released an AI-assisted animated short film titled Masha. The film was selected for the AI Frontier section of the Bucheon International Fantastic Film Festival (BIFAN), where it was screened and followed by audience Q&A sessions.

I am currently developing my next independent animated project while continuing to explore how I can combine my experience in traditional CG production with emerging AI technologies in the creative process. Find more of me at my personal website, my Masha short film, SIGGRAPH 2025 talk, and this SIGGRAPH 2026 poster.

What recurring production problem inspired this multi-layered RBD workflow, and why is revising a traditional destruction simulation after its overall motion has been approved so costly?

Jae Jun Yi: One of the main reasons I started developing this workflow was that art direction tends to become increasingly specific as a project progresses.

In the early stages, the overall direction of the destruction and the large-scale motion are usually the primary concerns. As the work develops, however, directors often begin to give much more detailed notes about the motion of individual pieces, the exact timing of a fracture, the size and shape of fragments, or a specific area of the frame.

At this stage, the goal is not necessarily to create the most physically accurate simulation possible. More often, the priority is how precisely the simulation can achieve the intended visual result on screen.

A traditional destruction workflow generally follows a process such as:
Fracture → RBD Simulation → Particle/Debris → Volume/Smoke
In production, however, several recurring problems can arise.

As art direction evolves, artists often need to recalculate a significant portion of the existing simulation, and in many cases the entire simulation has to be rerun. In an RBD simulation, many pieces physically interact with one another, so even a small change to one condition can alter the motion of other pieces that have already been approved.

For example, even if the overall motion has already been approved and you only want to add smaller fracture detail to one specific area, it is not simply a matter of increasing geometric detail. The number of pieces and their indexing change, and the constraint relationships may also need to be rebuilt, which means the connection to the existing simulation has to be reorganized as well.

That led me to think about destruction in a different way. Instead of trying to build the final level of detail from the beginning, I thought it could be more efficient to start with larger pieces and the overall motion, and then progressively fracture only the necessary areas into smaller pieces as the project and the art direction become more specific.

In other words, the level of simulation detail can also evolve progressively alongside the art direction. The main technical challenge is that each new simulation layer still has to connect naturally to the previous one. Because of that, an important part of the workflow is managing, on a per-piece basis, when a piece should continue following the existing motion and when it should transition into a new dynamic simulation.

Could you walk us through the layered simulation process? How is the motion from the approved base simulation preserved and passed into subsequent layers that introduce additional fractures, debris, or secondary effects?

Jae Jun Yi: The process begins with a standard first-pass RBD simulation.

At this stage, the main focus is not on detailed fracture patterns, but rather on the overall destruction motion and the composition within the frame. Once the base simulation has been sufficiently approved, I select the pieces that need additional detail.

There are two main ways to select target pieces.

One is to select them manually based on art direction, and the other is to select them procedurally using different conditions.

For example, target pieces can be selected based on:

• Pieces above a certain size
• Pieces with high velocity
• Pieces that receive a strong impact
• Pieces that collide with a specific object
• Pieces that travel beyond a certain distance

The selected pieces are then fractured again to generate smaller and more detailed geometry.

At this point, one of the most important things is to preserve the relationship between the pieces in the previous simulation and the newly fractured pieces.

Before creating the new fracture layer, I store the piece identifier from the previous stage as a separate attribute, and then generate a new indexing structure for the new fracture layer. This allows me to keep track of which original piece each new fragment came from.

Next, I transfer the transform and motion attributes from the previous simulation onto the newly created pieces. This means that creating a new fracture does not immediately start a new simulation.

Instead, each new piece first continues to follow the motion of the existing simulation. When a piece reaches its assigned trigger frame, it transitions from animated motion into an active RBD simulation. The trigger frame can be set manually by the artist, or it can be generated based on conditions such as impact or velocity.

The newly fractured pieces inherit the original motion through the piece identifier from the initial simulation, follow that motion until the specified trigger condition is reached, and then transition into active simulation.

Constraints are also transformed using the piece relationships from the previous simulation. When the new simulation becomes active, differences between the inherited motion and the newly simulated motion can sometimes cause constraints to stretch excessively. In those cases, I remove those constraints based on conditions such as constraint length.
After that, the new RBD motion can be used as a basis for secondary effects such as particles, debris, and dust.

If necessary, the same process can be repeated for a second, third, or additional simulation layer, allowing detail to be added progressively.

How do you manage collisions and physical continuity between layers so newly simulated fragments feel connected to the original destruction rather than appearing like separate effects placed on top?

Jae Jun Yi: I think this is the most important part of the workflow.

For a new simulation layer to feel like a natural continuation of the previous destruction rather than a separate effect placed on top, motion continuity between the layers has to be preserved.

To achieve this, every piece in the new simulation inherits the motion of the previous simulation. Even after a piece has been fractured again, it continues to follow the position, orientation, velocity, and other motion attributes of its parent piece until it becomes active.

Only when it reaches the specified trigger frame does it transition into active simulation. Even at that point, the simulation does not begin from an entirely new state. Instead, the inherited motion from the previous simulation is used as the initial state.

From that point onward, physical forces such as gravity, collisions, and constraints begin to affect that inherited motion.

Put simply, the new simulation does not replace the previous simulation. It continues from the previous motion and allows that motion to develop physically from that point forward. The same principle is applied again in the second and third simulation layers.

As a result, even though the final destruction may technically be built from multiple simulation passes, it can still appear as if the entire process is taking place within one continuous physical simulation.

What controls does the workflow give FX artists for art direction? Can they selectively refine individual regions, change fracture scale, or adjust timing without disturbing portions of the simulation that have already been approved?

Jae Jun Yi: This is a very important question because the starting point of this research was really about how FX artists can implement art direction more efficiently.
As a project progresses, the notes that FX artists receive tend to become increasingly specific.

For example:

“Keep the motion in this area exactly as it is.”

“I want the right side to break apart a little more.”

“I want smaller fragments to appear after this impact.”

“I want this piece to fall slightly later.”

In situations like these, rerunning the entire simulation can be very inefficient. So the first step in this workflow is to define what part of the simulation actually needs to be changed.

The most direct method is for the artist to manually select the target pieces.
If the director’s note refers to a specific region of the frame or a specific piece, manual selection can often be the fastest and most effective approach.

On the other hand, if the target pieces share a certain motion characteristic, procedural conditions can be used. For example, target pieces can be grouped based on: scale, velocity, impact, collision, and other conditions.

Another aspect I considered particularly important is the time range of the change. An approved simulation does not necessarily refer only to specific pieces. Sometimes the approved portion is defined in time.

For example, a director might say: “I like the motion up to this point, but I want it to break apart more after this frame.”

In that case, the existing motion can be preserved up to a certain point, and the new fracture simulation can be activated only after the required timing.
This gives the artist selective control over which pieces to modify, which region to refine, what fracture scale to use, and when the new simulation should become active.

I think this is one of the main advantages of the multi-layered workflow.

Which tools formed the foundation of the workflow, and did you need to develop custom Houdini setups, data structures, caching methods, or procedural tools to make layered revisions practical in production?

Jae Jun Yi: The workflow is fundamentally based on Houdini’s standard RBD simulation workflow.

Whenever possible, I prefer to extend an existing Houdini RBD workflow rather than build a completely separate system. To select target pieces and calculate activation conditions, I used a SOP Solver inside the RBD Solver.

For the conditions, I used information provided by Houdini’s RBD simulation, such as: velocity, impact, and collision data.

I also added several custom attributes to preserve information that needs to pass between simulation layers. For example, I store information such as the piece identifier from the previous fracture stage, target states, activation timing, and trigger values so that they can be used again in the next layer.

At the same time, I personally prefer to keep the number of nodes as low as possible and avoid moving too far away from a standard Houdini workflow.
There are two main reasons for this.

First, it makes the setup easier to modify. In FX production, art direction can change constantly, and the more specialized and complex a setup becomes, the more time it takes to modify and debug. Second, the setup needs to be understandable and usable by other artists. In a production environment, I think it is more important to maintain a structure that is familiar to Houdini artists than to build a highly complex system that only one artist can understand.

For that reason, each layer in the Multi-Layered RBD workflow does not use an entirely different system. Instead, the same fundamental structure is repeated and applied to each additional layer.

What are the main limitations of the approach, particularly for highly interconnected collapses or changes that fundamentally alter the base motion? Could similar layered techniques eventually improve destruction workflows for real-time cinematics or games?

Jae Jun Yi: The approach I proposed also has clear limitations. One of the most practical issues occurs when a new fracture layer becomes active and overlapping geometry or strong collisions occur between the existing pieces and the newly created fragments.

Because smaller pieces are being activated inside an already moving simulation state, some fragments may begin too close to one another or may intersect.

In those cases, the collision solver can react very strongly.
This can cause pieces to suddenly pop outward, vibrate, or move in unexpected directions. One practical solution I used was to slightly reduce the collision size of the pieces used in the new layered simulation. This creates a small amount of space between the fragments and helps reduce excessive collision responses at the moment of activation.

In the final result, these small differences are often difficult to notice because the RBD simulation is combined with secondary effects such as particles, debris, dust, and smoke.

However, geometry intersection still remains one of the limitations of the approach. Another clear limitation appears when the requested change is not a local detail adjustment, but instead changes the overall direction of the destruction or the large-scale motion itself. This workflow is optimized for adding detail while preserving an approved base motion. If the overall collapse behavior needs to change, it is generally more appropriate to rerun the base simulation.

For games and real-time applications, I think a similar approach could potentially be useful for baked destruction or cinematic sequences. In those cases, the idea is similar to film production because additional refinement is being added on top of motion that has already been established. However, the current approach has limitations for fully interactive destruction where the result needs to change dynamically in response to player input.

This workflow was primarily designed for film production, where the camera and the final screen composition are predetermined and where the director can give very specific art direction for the final frame. So the goal of this approach is not to create the most physically accurate simulation for every possible situation. Rather, it is to preserve approved motion as much as possible while giving FX artists more precise control over how they implement the director’s art direction for a specific composition and timing.

Jae Jun Yi, Senior FX Artist