Safety and Security

Field	Response
Model Application Field(s):	Media & Entertainment; Industrial/Machinery and Robotics; Healthcare (Biomechanics Research); Computer Vision Research; Animation and Simulation
Describe the life critical impact (if present).	None; this should not operate in safety-critical control loops (autonomous vehicles, medical devices, industrial safety systems).
Use Case Restrictions:	Abide by the Apache 2.0 License. SOMA must not be used to: (1) impersonate specific real individuals without their explicit consent; (2) generate synthetic body data intended to deceive biometric identification systems; (3) produce outputs that violate applicable laws or regulations in the deployment jurisdiction. Integration into safety-critical systems (medical devices, autonomous vehicles, industrial machinery) requires additional validation by the integrating team.
Model and dataset restrictions:	The Principle of Least Privilege (PoLP) is applied, limiting access for dataset generation and model development. Dataset access restrictions were enforced during PCA fitting and corrective MLP training. The released model artifacts (PCA components, skinning weights, rig data) do not contain raw scan data; access to the underlying SizeUSA dataset remains restricted to authorized NVIDIA personnel under the commercial license agreement. Bones RigPlay motion capture data (used for optional corrective MLP training) is commercially licensed (purchased by NVIDIA) and contains no real-person video or personally identifiable information; all motion sequences were retargeted to a fixed skeleton.
Security considerations:	SOMA processes numerical tensors (shape coefficients and pose parameters) only; it does not accept image, video, text, or executable inputs, substantially limiting its attack surface. The model does not make network calls at inference time. The NVIDIA Warp custom kernel is compiled and linked at package installation; users should verify package integrity via the official distribution channel (GitHub / Hugging Face). Report security vulnerabilities to NVIDIA here.
Description of methods to address potentially harmful data in training data:	The body scan datasets used for shape PCA fitting contain 3D geometric data only — no images, text, audio, or personally identifiable information. No screening for harmful content (CSAM, NCII, hate speech) is applicable to abstract 3D body geometry data. The GarmentMeasurement distillation data is synthetically generated using their source code and contains no personal or harmful content. Bones RigPlay motion capture sequences (used for optional corrective MLP training) were retargeted to a fixed skeleton prior to use, removing performer-specific biometrics; the dataset contains no visual appearance data or personally identifiable information.
Responsible AI practices:	SOMA is designed to represent diverse human body shapes; misuse for stereotyping, body shaming, or generating distorted body shapes to mock or demean individuals is contrary to the intended use. The ANNY backend is specifically recommended for applications requiring age-spanning (infant and children) or globally diverse body shapes to avoid reinforcing demographic biases from scan-collected datasets. Developers are responsible for implementing appropriate content guardrails in any user-facing application built on top of SOMA.