cs.CVMar 17, 2026
SaveSOMA: Unifying Parametric Human Body Models
Organizations: NVIDIA
Abstract
Parametric human body models are foundational to human reconstruction, animation, and simulation, yet they remain mutually incompatible: SMPL, SMPL-X, MHR, Anny, and related models each diverge in mesh topology, skeletal structure, shape parameterization, and unit convention, making it impractical to exploit their complementary strengths within a single pipeline. We present SOMA, a unified body layer that bridges these heterogeneous representations through three abstraction layers. Mesh topology abstraction maps any source model's identity to a shared canonical mesh in constant time per vertex. Skeletal abstraction recovers a full set of identity-adapted joint transforms from any body shape, whether in rest pose or an arbitrary posed configuration, in a single closed-form pass, with no iterative optimization or per-model training. Pose abstraction inverts the skinning pipeline to recover unified skeleton rotations directly from posed vertices of any supported model, enabling heterogeneous motion datasets to be consumed without custom retargeting. Together, these layers reduce the per-pair adapter problem to single-backend connectors, letting practitioners freely mix identity sources and pose data at inference time. The entire pipeline is fully differentiable end-to-end and GPU-accelerated via NVIDIA-Warp.
Figures & tables
Figure 1: SOMA unifies five heterogeneous parametric body models (SOMA-Shape, MHR, SMPL-X, Anny, and GarmentMeasurements) under a single animation pipeline. (a) Unified Skeleton: despite originating from entirely different identity spaces, joint hierarchies, and mesh resolutions, all five models are driven by the same SOMA skeleton in an identical pose, with no model-specific retargeting. (b) Unified Pose Correctives: a single MLP correctives model trained once on the shared SOMA topology produces anatomically plausible pose-dependent deformations for all backends, mitigating standard LBS artifacts without per-model corrective learning. (c) Unified Mesh Topology: all identity models share the same mesh structure, enabling skinning weights, deformation priors, and correctives to transfer seamlessly across backends.
Figure 2: Overview of SOMA . SOMA decouples body identity from pose through three sequential layers. Identity Model Provider (left): any supported backend (SOMA-shape, Anny, MHR, SMPL/SMPL-X, or GarmentMeasurements) maps its own shape parameters to a rest-shape mesh in its native topology. Bridging Layer (middle): two abstraction steps canonicalize the source identity into a unified representation. Mesh Topology Abstraction transfers the rest shape to the shared SOMA topology via pre-computed barycentric coordinates; Skeletal Abstraction then fits the shared -joint SOMA rig to the transferred rest shape in a single closed-form pass, with no iterative optimization or per-identity training. Animation Layer (right): all identity models are animated through the shared SOMA skeleton using joint rotations. When motion data arrives in another model’s convention, \ie , Pose Abstraction converts it to by analytically inverting the LBS pipeline; this step is bypassed when pose is already in the SOMA convention. A shared MLP Pose Correctives model then predicts pose-dependent vertex displacements to correct LBS artifacts, and Linear Blend Skinning produces the final posed mesh. The entire pipeline is fully differentiable end-to-end.
Figure 3: Training data for the SOMA -Shape PCA model. (a) A subset of the 9,326 SizeUSA body scans registered to the SOMA topology, exhibiting a wide range of body weights and proportions. (b) A subset of the 303 Triplegangers scans, registered to the same topology. All meshes are reposed to a canonical A-pose before PCA fitting, and mirror-augmented across the sagittal plane to enforce bilateral symmetry.
Figure 4: Mesh topology abstraction. Top: native mesh topologies of each identity model. Bottom: the same identities mapped to the shared SOMA topology via 3D barycentric interpolation. This common mesh serves as the pivot for all cross-model operations—skeleton fitting, pose transfer, correctives, and shape-space comparison all operate on a single canonical topology regardless of the source model.
Figure 5: Skeleton fitting on posed SAM 3D Body identities. Eight MHR identities in diverse poses with the SOMA skeleton fitted via SkeletonTransfer ( Sec. 3.4 ). Unlike joint regressors that assume a rest pose, our method generalizes to arbitrary posed shapes: joint positions are regressed via RBF interpolation, and joint rotations are recovered by Procrustes alignment, both in a single analytical forward pass with no iterative optimization.
Figure 6: Unified pose correctives across identity backends. Each row shows a different pose; columns correspond to SOMA -Shape, MHR, SMPL, Anny, and GarmentMeasurements. For each cell, the left mesh shows the corrective displacement applied to the canonical rest shape, and the right mesh shows the final posed result. A single correctives model trained once on SOMA ’s canonical topology produces anatomically plausible deformations for all backends.
| Backend | Src. Verts | Mean (mm) | Std (mm) | P95 (mm) | Wrap Mean | Wrap P95 |
| SOMA native | – | 0.0 | 0.0 | 0.0 | – | – |
| SMPL | 6,890 | 0.12 | 0.45 | 0.71 | 0.12 | 0.74 |
| SMPL-X | 10,475 | 0.06 | 0.22 | 0.45 | 0.06 | 0.47 |
| Anny | 13,718 | 0.01 | 0.12 | 0.01 | 0.01 | 0.01 |
| MHR | 18k | 0.40 | 0.73 | 1.49 | 0.31 | 1.28 |
Table 1: Topology abstraction fidelity across backends. Closest-point-to-mesh distance (mm) between the SOMA rest shape and the native source mesh, measured over 100 diverse identities per backend. “Wrap” is the baseline registration error of the pre-registered SOMA wrap mesh against the source neutral mesh. Vertices in facial inner geometry (eye bags, mouth bag) and between-toes regions—which have no correspondence in most source topologies—are excluded.
Figure 7: Shape diversity across backends driven by a single pose. Three sampled identities per backend— SOMA -Shape (green), MHR (blue), SMPL (pink), and Anny (yellow)—all driven by the same skeletal pose. Despite originating from entirely different generative models, all bodies share the same pose interpretation, illustrating the plug-and-play nature of SOMA identity-pose decoupling.
| Method | Mean (mm) | Median (mm) | Max (mm) | FPS |
| Skel. transfer only | 16.5 | 13.9 | 80.1 | 17,393 |
| Analytical | 5.3 | 3.2 | 88.1 | 882 |
| Autograd FK (no init) | 501.8 | 479.4 | 1354.2 | 79 |
| Autograd FK (w/ init) | 4.1 | 2.1 | 81.5 | 78 |
| Analytical + Autograd (10) | 7.8 | 6.4 | 88.6 | 435 |
Table 2: Pose inversion accuracy and throughput on AMASS. Per-vertex reconstruction error (mm) and throughput (frames/sec) on an NVIDIA A100 GPU. “Skel. transfer” is the raw skeleton-fitting initialization with no iterative refinement. “Analytical” adds inverse-LBS refinement with Newton-Schulz orthogonalization (body=2, full=1). “Autograd FK” optimizes 6D rotation parameters through FK+LBS with Adam (100 iterations); “no init” starts from the bind pose, “w/ init” warm-starts from the skeleton transfer result. “Analytical + Autograd” warm-starts autograd from the analytical result (10 iterations).
| Method | All | Body | Hands | Feet | Head |
| Analytical | 8.8 | 16.8 | 4.7 | 8.2 | 6.9 |
| + Autograd (100) | 6.6 | 15.3 | 2.0 | 5.8 | 4.8 |
| Reduction | 25% | 9% | 57% | 29% | 30% |
Table 3: Per-region pose inversion error (MHR, 200 SAM 3D Body frames). Mean per-vertex error (mm) by body region. “Analytical” = body=2, full=1. “+ Autograd” adds 100 Adam iterations warm-started from the analytical result.
Figure 8: Hand zoom: analytical vs. autograd FK refinement. MHR backend on SAM 3D Body (teal = ground truth, red = SOMA reconstruction). (a) The analytical solver is near-optimal globally but leaves residual misalignment at the fingertips. (b) Adding 100 autograd FK iterations redistributes error toward the body trunk (visible as a slightly increased offset in the background), achieving a tight overlay at the fingers.
Figure 9: Pose inversion quality across methods. Top row: SMPL backend (yellow = ground truth, purple = SOMA reconstruction). Bottom row: MHR backend on SAM 3D Body (teal = ground truth, red = SOMA reconstruction). (a, d) Skeleton transfer alone provides a coarse estimate with visible misalignment at the extremities. (b, e) Analytical refinement with Newton-Schulz orthogonalization closely tracks the ground truth. (c, f) Adding autograd FK refinement achieves the tightest overlay, reducing residual error at the feet and hands.
Figure 10: Pose inversion ablations. (a, b) Shoulder zoom from a single SMPL frame (yellow = GT, purple = SOMA ). SVD-based Kabsch alignment exhibits shoulder popping due to a singular-vector sign flip; Newton-Schulz produces a smooth result. (c, d) MHR on SAM 3D Body (teal = GT, red = SOMA ). Without initialization, 100 autograd iterations converge to an incorrect local minimum; the analytical solver recovers the pose correctly.
| Mode | Batch | Skel. (ms) | Total (ms) | Meshes/sec |
| Warp (GPU) | 1 | 0.8 | 2.1 | 476 |
| Warp (GPU) | 8 | 0.9 | 3.4 | 2,353 |
| Warp (GPU) | 32 | 1.1 | 6.8 | 4,706 |
| Warp (GPU) | 128 | 1.4 | 18.2 | 7,033 |
| PyTorch (CPU) | 1 | 3.2 | 12.1 | 83 |
| PyTorch (CPU) | 8 | 4.1 | 38.7 | 207 |
Table 4: Runtime performance. Throughput (meshes/sec) and per-call latency (ms) of the SOMA forward pass. Breakdown shows skeleton fitting cost (RBF regression + Kabsch) vs. total forward pass. Warp = GPU path (NVIDIA Warp LBS kernel); PyTorch = CPU dense path. Identity backend: SOMA -native (no topology abstraction overhead).
| Model | Mean (mm) | Median (mm) | P95 (mm) | |
| SMPL | 10 | 14.11 | 12.31 | 30.49 |
| GarmentMeasurements | 15 | 11.81 | 10.67 | 24.18 |
| SOMA -Shape (Ours) | 128 | 5.82 | 4.81 | 13.60 |
| SMPL-X | 300 | 5.45 | 4.34 | 12.97 |
Table 5: Cross-model PCA reconstruction on 33 held-out body scans. Per-vertex distance (mm) between the unposed scan and the PCA reconstruction at each model’s full component count . All models are evaluated through the same SOMA pipeline, differing only in the target topology, rest pose, and PCA basis.
Figure 11: Cross-model PCA reconstruction error on held-out body scans. SOMA ’s topology and pose abstraction layers enable a principled cross-model comparison. Each column shows a different shape model’s mesh topology, all rendered in SOMA A-pose; each row (a–d) is the same identity. Color encodes per-vertex error (blue = 0 mm, white = 10 mm, red 20 mm). SMPL (10 components) and GM (GarmentMeasurements, 15) show widespread residual, while SOMA -Shape (128) and SMPL-X (300) achieve low error across the body surface.