NVIDIA releases SOMA-X v0.3.0 unified parametric human body model framework
nvidia/SOMA-X
NVIDIA has released SOMA-X v0.3.0, a unified framework for parametric human body models that decouples identity representation from pose parameterization. The release supports six identity backends including SOMA-shape, SMPL, and SMPL-X, and introduces native hand layers with interoperability for user-supplied MANO models. Designed for commercial use under the Apache 2.0 license, the model enables real-time simulation pipelines on NVIDIA GPUs with throughput exceeding 7,000 meshes per second on an A100. The core pipeline relies on closed-form analytical modules rather than learned neural networks, though an optional shallow MLP can be enabled for surface correction.
NVIDIA releases SOMA-X v0.3.0, an Apache 2.0 licensed analytical framework that decouples human body identity from pose parameterization to enable real-time simulation and animation across six backends (SOMA-shape, SMPL, SMPL-X, MHR, ANNY, GarmentMeasurements) with optional hand layers.
Source: NVIDIA — Robotics model repository updates · Read original article ↗
English translation unavailable. Choose the original article to read it.
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
meshes per second
>7,033
View original evidence
of model parameters: Core analytical pipeline: 0 learned parameters (closed-form) Principal Component Analysis (PCA) for SOMA-shape model: 128 principal components x ~18,000 mid-LOD vertices x 3 = ~3.1 x 10^6 coefficients (pre-fitted, not gradient-trained) Optional pose corrective Multilayer Perceptron (MLP) layers: ~1 × 10^8 parameters (if enabled) Computational Load Throughput: > 7,033 posed meshes/second on NVIDIA
Open source S9meshes per second
7,000
View original evidence
eal-time throughput exceeding 7,000 meshes per second at batch size 128 on an A100 GPU. Software Integration: Runtime Engine(s): NVIDIA Warp (GPU-accelerated LBS kernel, torch.export -compatible) PyTorch (CPU and GPU fallback) N/A — No dependency on TAO, Riva, NeMo, or other NVIDIA SDK runtimes Supported Hardware Microarchitecture Compatibility: NVIDIA Ampere (A100, A30, A40, A10, RTX 3000-series) — tested on A100 80
Open source S16latency per mesh (GPU)
2.1
View original evidence
A100 80GB (batch size 128, GPU Warp path) Latency: 2.1 ms per mesh (batch = 1, GPU); 12.1 ms (batch = 1, CPU 32-core) Skeleton fitting: < 1.68 ms (batch = 1) Training compute: N/A — core pipeline requires no gradient-based training; PCA shape space fitted offline from body scan data. Input(s): Input Type(s): Numerical tensors (floating-point) Input Format(s): Identity coefficients: floating-point tensor, shape (B, K
Open source S10latency per mesh (CPU)
12.1
View original evidence
A100 80GB (batch size 128, GPU Warp path) Latency: 2.1 ms per mesh (batch = 1, GPU); 12.1 ms (batch = 1, CPU 32-core) Skeleton fitting: < 1.68 ms (batch = 1) Training compute: N/A — core pipeline requires no gradient-based training; PCA shape space fitted offline from body scan data. Input(s): Input Type(s): Numerical tensors (floating-point) Input Format(s): Identity coefficients: floating-point tensor, shape (B, K
Open source S10skeleton fitting latency
<1.68
View original evidence
A100 80GB (batch size 128, GPU Warp path) Latency: 2.1 ms per mesh (batch = 1, GPU); 12.1 ms (batch = 1, CPU 32-core) Skeleton fitting: < 1.68 ms (batch = 1) Training compute: N/A — core pipeline requires no gradient-based training; PCA shape space fitted offline from body scan data. Input(s): Input Type(s): Numerical tensors (floating-point) Input Format(s): Identity coefficients: floating-point tensor, shape (B, K
Open source S10meshes per second (batch 128)
>7,033
View original evidence
of model parameters: Core analytical pipeline: 0 learned parameters (closed-form) Principal Component Analysis (PCA) for SOMA-shape model: 128 principal components x ~18,000 mid-LOD vertices x 3 = ~3.1 x 10^6 coefficients (pre-fitted, not gradient-trained) Optional pose corrective Multilayer Perceptron (MLP) layers: ~1 × 10^8 parameters (if enabled) Computational Load Throughput: > 7,033 posed meshes/second on NVIDIA
Open source S9meshes per second (batch 128)
7,000
View original evidence
eal-time throughput exceeding 7,000 meshes per second at batch size 128 on an A100 GPU. Software Integration: Runtime Engine(s): NVIDIA Warp (GPU-accelerated LBS kernel, torch.export -compatible) PyTorch (CPU and GPU fallback) N/A — No dependency on TAO, Riva, NeMo, or other NVIDIA SDK runtimes Supported Hardware Microarchitecture Compatibility: NVIDIA Ampere (A100, A30, A40, A10, RTX 3000-series) — tested on A100 80
Open source S16
Dataset details reported by the source
- License
- commercially licensed Open source S21
- Version
- v0.3.0 Open source S5
- Modalities
- 3D meshes and 3D motion data Open source S20
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: not_reported
What remains unknown
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
h) pairs. Data Collection Method: Automatic/Sensors (structured-light 3D body scanners) for SizeUSA and TripleGangers; Synthetic for GarmentMeasurement distillation; motion capture for Bones RigPlay. Labeling Method: Automatic/Sensors (body landmarks auto-detected from scans) Dataset License(s): SizeUSA — commercially licensed (purchased by NVIDIA); TripleGangers — commercially licensed (purchased by NVIDIA); Garment
Open source S21
del — body shape data distilled from the GarmentMeasurement parametric model to augment the shape space with garment-relevant proportions. Shallow MLP for Pose Correctives: Bones RigPlay Dataset : 80,000 (pose,mesh) pairs samples from Bones RigPlay motion capture dataset owned by NVIDIA. Data Modality Other: 3D meshes and 3D motion data Non-Audio, Image, Text Training Data Size : 20,000 3D meshes and 80,000 (pose,mes
Open source S20
rgest anthropometric survey of the U.S. population, covering diverse body shapes across age, sex, and BMI groups. Used to compute the 128-component PCA shape space for the SOMA-native identity backend. TripleGangers — commercially licensed 3D body scan dataset purchased from TripleGangers, containing body scans of 303 individuals. Contributes additional shape diversity to the SOMA-shape PCA. GarmentMeasurement PCA mo
Open source S19
ans — freely mix identity sources with SOMA's pose representation. Simulation and robotics — lightweight analytical forward pass enables real-time simulation pipelines with diverse body shapes. Hand animation and reconstruction — wrist-local left/right hand layers support mid, low, and extra-low LODs and a sampled articulation prior. Release: SOMA-X v0.3.0 Reference(s): SOMA: Unifying Parametric Human Body Models. SM
Open source S5
of model parameters: Core analytical pipeline: 0 learned parameters (closed-form) Principal Component Analysis (PCA) for SOMA-shape model: 128 principal components x ~18,000 mid-LOD vertices x 3 = ~3.1 x 10^6 coefficients (pre-fitted, not gradient-trained) Optional pose corrective Multilayer Perceptron (MLP) layers: ~1 × 10^8 parameters (if enabled) Computational Load Throughput: > 7,033 posed meshes/second on NVIDIA
Open source S9
eal-time throughput exceeding 7,000 meshes per second at batch size 128 on an A100 GPU. Software Integration: Runtime Engine(s): NVIDIA Warp (GPU-accelerated LBS kernel, torch.export -compatible) PyTorch (CPU and GPU fallback) N/A — No dependency on TAO, Riva, NeMo, or other NVIDIA SDK runtimes Supported Hardware Microarchitecture Compatibility: NVIDIA Ampere (A100, A30, A40, A10, RTX 3000-series) — tested on A100 80
Open source S16
A100 80GB (batch size 128, GPU Warp path) Latency: 2.1 ms per mesh (batch = 1, GPU); 12.1 ms (batch = 1, CPU 32-core) Skeleton fitting: < 1.68 ms (batch = 1) Training compute: N/A — core pipeline requires no gradient-based training; PCA shape space fitted offline from body scan data. Input(s): Input Type(s): Numerical tensors (floating-point) Input Format(s): Identity coefficients: floating-point tensor, shape (B, K
Open source S10
Implications for data suppliers
RoboSignal interpretation and collection questions, not statements of buyer demand.
- Confirm the required data type and collection setting with the buyer; this source does not establish a complete collection specification.
- Compare the reported units and scope before using these quantities in a budget. Recording hours, sensor-hours and trajectories are different measures.
- Check the exact dataset version, license, modalities and task coverage against the intended use.
- Validate demand and acceptance criteria with a buyer before scaling. Publication, popularity and a research result do not establish a purchase commitment.
Source：NVIDIA — Robotics model repository updates · huggingface.co