ParaHuman — SOMA-X Body Runtime

Drive the SOMA-X parametric body model natively in Unity. Full pose / shape / bone-proportion control, Burst-powered neural correctives, a two-level LOD body with one-click setup.

ParaHuman brings NVIDIA Research's SOMA-X parametric human body to your Unity project.

SOMA-X is a high-fidelity parametric human body model released by NVIDIA Research: a 77-joint skeleton with 32 procedural twist joints, a 128-coefficient body-shape space, per-bone length ratios and a neural pose-corrective network. This toolkit runs the SOMA-X body model entirely inside Unity: no Python, no external runtime, no network calls.

Exact, verified reproduction Skeleton articulation, joint orientation, procedural twist joints, the body-shape pipeline and the pose-corrective network are reimplemented natively in C# and Burst, and numerically validated against the official model to within 0.01 mm. Parameter layout and conventions match SOMALayer.pose(poses, transl) of the official package exactly.

Neural pose correctives, real-time SOMA-X's pose-dependent deformations are evaluated in Burst-compiled parallel jobs, ~1–2 ms per body, with all bodies in the scene scheduled together so the work overlaps across CPU worker threads. Each LOD renderer gets its own corrective set on a per-body mesh instance.

Body shape and proportions

• SetIdentity(float[128]) — PCA body shape: builds a new rest mesh, refits the skeleton and updates bindposes and bone lengths on every LOD renderer (~20 ms)
• SetBoneScales(float[60]) — bone-length ratios, independent of shape
• 32 procedural twist joints driven automatically from the pose

Complete parameter control

• SetPose(float[77*3], hipsTranslation) — axis-angle per joint, T-pose relative
• SetPoseRotations(Quaternion[77], hipsTranslation)
• Inspector panel with shape sliders, pose sliders grouped by body region and the Hips translation — live in Edit mode, values serialized with the scene

Bring your own SOMA data Because the parameter layout matches the official model exactly, parameters produced by any SOMA-based system can be fed straight into the API for display. Includes a motion-clip player (poses + Hips translation, optional in-place mode, optional per-clip body shape) and four sample clips: the official SOMA-X example animation (931 frames) and three sequences derived from stock footage with an external reconstruction tool. The clip format is documented in the README, so you can convert your own motion data; no converter tools are included.

Note: this asset is a rendering and animation runtime. It does not perform motion capture, pose estimation or video reconstruction — pose/shape parameters come from your code, the included clips, or external tools (none of which are included).

Two LOD levels, one click A mid mesh (18k vertices) and a low mesh (4.5k vertices), each with matched corrective data. The included builders (Tools > SOMA > Build Body... and Build LOD Body...) assemble a body with one skeleton, a SkinnedMeshRenderer per LOD and a native Unity LODGroup, with pose correctives, body-shape support and the parameter controller wired up as options.

What's included

• Full C# source (runtime + editor), assembly definitions, LMotion.SOMA namespace
• Skeleton, twist and bone-scale tables; mid and low meshes; corrective data for both LODs; the 128-coefficient body-shape model
• Ready-made body prefabs, body builders, parameter inspector, motion player
• 4 demo scenes: parameter control, LOD switching, LOD comparison, motion playback
• Four motion clips, documentation, changelog and third-party notices

Requirements

• Unity 6000.x with the Universal Render Pipeline (developed on URP 17.5) — the demo scenes and the included ground material are URP. The runtime itself is pipeline-agnostic: it drives transforms, mesh vertices and skinning data and never touches shaders, so the body renders in any pipeline once you assign your own material
• Burst + Collections packages (free, via Package Manager)
• Quality Settings: Skin Weights = Unlimited (the mesh uses up to 10 bone influences per vertex; the toolkit warns about misconfiguration automatically)

Use cases Visualizing externally reconstructed poses · synthetic training-data generation · crowds and NPCs · avatar and character-creator systems · biomechanics and research demos · fitness / health applications

This asset is an independent product and is not affiliated with, sponsored, or endorsed by NVIDIA Corporation. "SOMA", "SOMA-X" and "GEM-X" are used solely to describe compatibility and the provenance of included data. SOMA-X model data is redistributed under the Apache License 2.0; the sample motion clips were produced from Pexels stock footage with external tools and contain only numeric body-model parameters — full notices included in the package.