Love Handles: Decimation for Deformation Handles with Compact Support and Low Memory Footprints
David IW Levin, Paul Kry, Kartic Subr, Ryan Schmidt, Etienne Vouga, Teseo Schneider
cs.GR
2026-08-18
The first decimation-based algorithm computes sparse, compactly supported deformation handles: a 797K-tetrahedra metamaterial basketball compresses to 268 handles at 5% error, enabling nonlinear elastodynamics at 150 steps per second on a DGX Spark.
Soft-body simulation for animation, games, and robotics scales poorly with mesh resolution. A 200K-vertex metamaterial basketball solved at every step in full space is nowhere near real time. Reduced-order methods (ROMs) attack this by driving motion with a few hundred handles (affine transforms that control groups of vertices, the same formalism as character skinning) instead of hundreds of thousands of free vertices.
Generating handles and weights automatically is the open problem. Existing routes either assume handle positions and fit weights post hoc, or cluster first and fit weights second. Both suffer from the same disease: eigenanalysis produces dense, globally supported weights, coupling distant parts of the object. Change the boundary conditions (the object lands, gets clamped) and the system locks into unnaturally stiff deformations. Dense weights also scale storage with every added handle, capping the expressivity of the reduced space. Sparsity regularizers help but do not guarantee compact support; clustering metrics undersample stiff-but-thin features like tree branches that absolutely should deform.
The idea is borrowed from classic 1997 mesh simplification (Garland-Heckbert edge collapse): start with far too many handles and delete the cheapest one at a time until a user-set error tolerance is met.
Error-driven removal also fixes the clustering failures: thin, stiff regions carry large target-field signal, so deleting their handles spikes the error and the algorithm keeps handles where they matter.
At runtime, a compact-support-aware reduced cubature scheme splits the mesh into disjoint elements by handle subset, solves NNLS per element to pick a handful of quadrature tetrahedra, and assembles in parallel on disjoint bases. The pipeline runs in NVIDIA Warp on a DGX Spark (Blackwell, 128GB unified memory), entirely in FP32.
| Setting | Method | Result |
| 797K-tet basketball (25 modes, ε=5%) | Ours | 268 handles, ARAP elastodynamics at 150 steps/s |
| Equal-memory accuracy, 7 models, k=25 | vs Brandt et al. 2018 | 6.04× lower error on average, up to 15.6× on complex meshes |
| Equal-memory accuracy, 7 models | vs Li & Barbič 2019 | 8.34× lower error on average, up to 20.28× |
| 15 meshes, ε=5% | scalar DOF reduction | 90%+ for all but the coarsest Bunny |
| Per-step time (error 0.05, 25 modes) | all models | rarely above 5 ms; PCG solve and floor contact dominate |
| Golden Gate Bridge interactive demo | simulation + rendering | 30 fps throughout, penalty-spring contact in real time |
The cantilever beam comparison (Fig. 3) is the sharpest evidence: with no boundary conditions baked into either basis, linear modes lock severely at equal memory and moderately at equal DOFs (10× the memory), while decimated handles bend correctly on contact.
This is a directly usable asset pipeline for games and robot learning. RL wants hundreds of simulation steps per second, and current robotics frameworks ship only rudimentary soft-body dynamics; paying 36 preprocessing hours once for an asset resimulated millions of times is a good trade. The fixed 6-handles-per-vertex budget gives predictable memory, which suits modern GPUs where compute is cheap and VRAM is not. The authors commit to releasing all code, data, handles, and cubature rules.
Calibration: this is not new reduction theory. It ports a decades-old simplification weapon to handle computation for the first time, and wins on clean engineering.