Tim McGraw, an associate professor at Purdue, and his student Jack Myers built a system that turns cryosection imagery — including Visible Human Project data — into virtual bodies you can cut, tear, retract, and separate in real time, on ordinary GPUs. It won the Audience Choice award at SIGGRAPH 2026’s Real-Time Live!. The interview is a walkthrough of one representation decision and everything that follows from it.

Existing anatomy tools already use this data, but only to clip it against a plane or make structures transparent. McGraw’s problem is interactivity: deforming and cutting the way you would in a dissection lab.

The build, in two passes:

  • A low-resolution grid over the cryosection image creates position-based dynamics (PBD) particles where the specimen is
  • A 4x higher-resolution grid creates volumetric splats over the same area
  • Each splat carries a signed distance value updated at runtime — its distance from the model’s visible surface
  • SDF intersection and subtraction do the fracturing; because the field updates as the model breaks, only the outer shell of splats is drawn (internal and back-facing splats are culled)

Why splats instead of a mesh:

  • Splats have no connectivity, so a cut is just reassigning splats from one region to another
  • Cutting a triangle or tetrahedron produces a shape that is neither, and the cut has to propagate to neighbours
  • That propagation is exactly the kind of mess that compute shaders handle badly

The hardest part wasn’t the math. It was decomposing the work into parallel dispatches without read/write conflicts — dispatch on the region grid, then on particles, then on regions again to clean up — and minimising how many of those separate dispatches are needed. That, McGraw says, still is the challenge.

Portability was a constraint, not an afterthought: everything runs in compute shaders, with no game engine and no CUDA, so the technique isn’t welded to one machine.

The realism trade-off is deliberate. This isn’t teaching biomechanics of tissue — it’s teaching the names, shapes, and relative positions of structures, in an environment where a wrong cut can be undone. McGraw leans on the game-physics argument: plenty of collision approximations go unnoticed, and the same should hold here.

What he won’t claim is that it works educationally. Educational value needs a pre-test, post-test, and a retention follow-up, which needs a system actually aimed at teaching rather than a demo of a technique. So the current work is talking to educators.

The obvious next home is games and VFX: small-to-medium destruction, soft-body breakage, creature dismemberment, pre-vis. Models can also come from voxelised meshes, not just medical scans. McGraw’s hope is that it gets used with the spirit of practical-effects gore in The Thing or Evil Dead — and not with real human cryosection data.

A useful reminder of a general pattern: pick the representation that makes your one hard operation cheap. Meshes make topology change expensive; splats make it bookkeeping. Everything else in the project is downstream of that choice.