OmniBody: KAIST’s Real-Time Soft-Body Simulation for Cutting and Fragmentation

OmniBody, a simulation framework for deforming, cutting, and fragmenting soft bodies, was developed by Korea Advanced Institute of Science and Technology (KAIST) researchers Jewoo Shin and Jinah Park. The work was presented as part of the Animation & Simulation posters at SIGGRAPH 2026; an interview in which Shin discussed the research was published on 12 October 2026. OmniBody’s key difference is that it manages connections between particles when an object is cut, rather than rebuilding its mesh structure.

Local simulation with ellipsoid particles

Instead of traditional meshes made up of connected vertices and faces, OmniBody uses voxel-aligned, oriented ellipsoid particles. Local Extended Position-Based Dynamics (XPBD) constraints between particles govern contact, bending, elasticity, volume preservation, and fracture behavior. In the researchers’ approach, the object’s overall motion emerges from the combination of these local interactions; the method does not require matrix assembly or solving a system of linear equations.

The ability to tune constraints independently makes it possible to create different material responses. For example, increasing the compliance of the contact constraint produces a more gel-like behavior, while volume-preservation settings change how much the object can stretch and compress. Shin says these parameters can be adjusted during a simulation.

How do cutting and tearing work?

During tearing, the connection between neighboring particles is permanently removed when the distance between their surfaces exceeds a threshold. Cutting is performed by disabling contact connections along the path followed by a virtual cutter. This is intended to avoid steps such as vertex duplication and remeshing, which may be required in mesh-based methods when topology changes.

However, the method is designed for general tearing and fragmentation rather than precisely following a specific crack line. According to Shin, the particles’ equal mass makes it difficult to guide a fracture along a desired path. Cleanly displaying the internal surfaces exposed by a cut is another visual challenge to address. The current visualization is based on drawing particles directly or using a mesh that tracks their motion; both options have limitations for fully fragmentable models.

Performance measurements and areas for improvement

In tests conducted by the researchers on an NVIDIA GeForce RTX 3080 Ti, the simulation time per frame was 0.9 milliseconds for an 8,000-particle 20³ grid and approximately 3.1 milliseconds for a 64,000-particle 40³ grid. In a more demanding test cutting a stretched 40³ cube, the peak frame time remained below 10 milliseconds. These are test results reported by the team; they do not mean the same timings will apply to different scenes and systems.

The team says the neighborhood structure’s division of constraints into three groups along the x, y, and z axes is well suited to GPU parallelization. OmniBody is still under development, however. Running many small operations sequentially and the memory layout used per particle are among the areas that could be improved. The researchers also note that the physical behavior does not always produce realistic results and that interactions between multiple objects need to be made more robust.

Implications for architecture and visualization workflows

Based on the information currently available, OmniBody is not a feature of an architectural modeling or visualization application; it is research into interactive, deformable objects. Architects, interior designers, and archviz teams should therefore view it not as a tool that can be added directly to a production pipeline today, but as an approach to simulating objects that can physically change in virtual environments. Teams developing real-time experiences or XR scenarios may want to follow potential applications for this type of interaction.

Before evaluating it in a production workflow, compatibility with the target software and engines, hardware requirements, and licensing terms should be verified; the source does not specify any integration or licensing information. How exposed surfaces are rendered after cutting can also affect visual quality. Shin does not suggest that OmniBody should replace existing soft-body or cloth methods in every use case; its advantages may be more limited for tasks where predefined fragmentation or mesh-cutting techniques are sufficient.