The Material Point Method (MPM) represents deformable material as a cloud of Lagrangian particles that carry mass, velocity and deformation, transferred through a background Eulerian grid each timestep (particle→grid→particle). It handles large deformation, splitting and merging for free — no remeshing, no explicit contact handling between particles of the same body.
The snow itself follows Stomakhin et al. 2013, "A Material Point Method for Snow Simulation" (Disney) — fixed-corotated elasticity plus a hardening plasticity model: the elastic part of the deformation gradient is clamped to a critical compression/stretch range, and the Lamé parameters harden exponentially with accumulated plastic compaction. That single rule is what gives packed snow its characteristic mix of springy give and permanent crunch.
The transfer scheme is MLS-MPM (Hu et al., SIGGRAPH 2018), simulated in 3D on an RTX 2070 with taichi. Precomputed offline, then quantized and streamed to this page for playback. The letters are solid snow blocks, extruded from a real typeface and dropped onto a rigid floor. What makes them shatter rather than splat is the tensile side of the plasticity: the critical stretch is set very low, so material pulled apart yields almost immediately and loses nearly all strength, letting cracks open and whole chunks separate — while the bulk, still under compression, stays stiff enough to keep its shape on the way down.
Generation script lives in this repo under gen/sim.py. There is a full technical report covering the solver, the surface-reconstruction pipeline, and the three bugs that made fracture impossible at first. See also the original paper and the authors' reference MPM code.