NVIDIA SOMA-X: One Open-Source Skeleton for Every Human Body
— by Vivax
On 18 June 2026, NVIDIA released SOMA-X v0.2 — an open-source, Apache-2.0 framework that can represent virtually any human body from a single canonical…
On 18 June 2026, NVIDIA's research lab released SOMA-X v0.2, an open-source human body model that can represent virtually any body type — tall or short, slender or broad, adult or child — from a single canonical skeleton. Published under a permissive Apache-2.0 license and ready for commercial use, SOMA (the framework behind SOMA-X) tackles a problem that has quietly held back 3D human modeling for years: the field's most useful body models do not speak the same language. SOMA-X is NVIDIA's attempt to make them interoperable, and the v0.2 update adds the deformation quality, rig assets and tooling needed to put it to work in real production pipelines.
To understand why this matters, you have to understand the mess it cleans up. Parametric human body models — SMPL, SMPL-X, MHR, Anny and others — are the foundation of human reconstruction, animation and simulation, yet each one uses a different mesh topology, skeleton, shape parameterization and unit convention. Combining them has traditionally meant writing bespoke converters for every pair of models — an O(M²) explosion of brittle, hand-maintained adapters. SOMA collapses that to O(M): each model connects once to a single shared representation, and from there everything can talk to everything else.
Structurally, SOMA decouples identity — what a body looks like — from pose — how it moves — and routes both through one canonical mesh and skeleton using three abstraction layers. A mesh-topology layer maps any source model's body onto a shared canonical mesh. A skeletal-abstraction layer fits identity-adapted joints into any body shape in a single closed-form pass, with no iterative optimization and no per-model training. And a pose-abstraction layer inverts the skinning math to recover the unified skeleton's rotations directly from posed vertices, so motion captured in SMPL or MHR can be replayed on the SOMA rig without custom retargeting. The whole pipeline is end-to-end differentiable and GPU-accelerated through NVIDIA Warp, built around a full-body rig of 77 joints.
The v0.2 release is what turns the research into something usable. It adds procedural twist joints for better deformation in articulated motion, an extra-low level-of-detail body for lightweight fitting, pose-conversion utilities across SOMA, SMPL and SMPL-X, a standard USD rig asset (SOMA_template_rig.usda) for 3D and DCC pipelines, and reference implementations for Blender and Maya — alongside public API documentation. Around it sits a fast-growing ecosystem: GEM-X recovers full 77-joint body, hand and face motion from a single video, a SOMA retargeter maps human motion onto humanoid robots, and large motion-capture datasets are now published directly on the shared SOMA skeleton.
We're paying close attention at Vivax. Faithful 3D human modeling underpins a great deal of what clinical AI needs — posture and gait analysis, rehabilitation tracking, surgical and procedural simulation, and patient-specific anatomy — and the cost of stitching together incompatible body models has always been real. An open, Apache-2.0 framework that unifies them on one differentiable skeleton is exactly the kind of building block we want to put to the test. We're looking forward to running SOMA-X as an A/B test against our current human-modeling implementations — measuring fidelity, speed, and how cleanly it fits into our privacy-first pipeline — and we'll share what we learn.