SOMA-X

SOMA-X

Maps several human body models onto one shared body, hand and skeleton layout so a single posing pipeline works with all of them. For vision, animation, simulation and robotics teams.

Made by nvidia on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: SOMA-X is a model from nvidia for human body shapes and poses, filed under robotics. Six body model families, SMPL and ANNY among them, are mapped onto a single shared body, hand and skeleton layout. The same posing code then works with any of them. The v0.3 release also includes left and right hand layers.

Who it suits: It suits people who build with 3D human bodies, such as vision and graphics researchers, animators, and teams doing robotics or simulation work.

Industry | Who uses it | How they use it | Example
Computer vision | Researchers | Estimate poses and rebuild human bodies from images | Swapping one body model for another without retraining the pose side
Animation | Graphics and animation engineers | Apply motion capture sequences to any supported body model | One axis-angle motion clip played on several different bodies
Robotics | Robotics researchers and companies | Run real time simulations with many body shapes | A simulation batch where every person has a different build
Machine learning | Machine learning engineers | Feed identity and pose tensors into training or inference code | Batches of body coefficients and joint rotations

You can skip it if you do not work with 3D human bodies.

Where to use it:

• Python, with inputs given as float32 PyTorch tensors
• NVIDIA GPUs through Warp kernels, or a CPU
• Real time simulation pipelines
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
Pose estimation and body reconstruction | Vision researchers | A body mesh you can swap between model families
Motion transfer | Animators | Capture sequences played on any supported body
Hand animation | Graphics engineers | Wrist local hand meshes at mid, low or extra low detail

How to set it up:

• Python and pip
• An NVIDIA GPU if you want the Warp path, since the card also times a CPU run
• Your own MANO models, only if you want them for the hands

1. Install the Hugging Face hub package.
2. Download the model files to your computer.
3. Pick the identity backend you want and prepare its coefficients as float32 tensors.

pip install -U huggingface_hub
hf download nvidia/SOMA-X

How to use it: You pass identity coefficients and pose parameters as tensors, with B as the batch size, and get posed meshes back. The table lists the input shapes.

Input | Shape
Identity coefficients | (B, K), with K = 128 for SOMA-shape
Pose as axis-angle | (B, 77, 3)
Pose as rotation matrices | (B, 77, 3, 3)
Optional root translation | (B, 3)
Hand pose as axis-angle | (B, 25, 3)
Hand identity coefficients | (B, 20)

The core pipeline has three steps, and the card describes each one.

Tool | What it does
Barycentric Mesh Transfer | Moves a body from one backend's mesh layout to the shared one with a single sparse matrix product
RBF Skeleton Fitting | Fits the 77 joint skeleton to the identity in one linear solve
Linear Blend Skinning | Bends the mesh to your pose, with a GPU path through NVIDIA Warp
Optional corrective MLP | Reduces surface artifacts that depend on the pose

Example: Take one motion capture clip in axis-angle form and play it on a SOMA-shape body, then on an SMPL-X body, using the same pose tensors both times. You get a posed mesh for each body without retraining anything.

For newcomers and experts: SOMA-X is not a chat model. It turns a list of body shape numbers and joint rotations into a posed 3D body. For a first try, download the files, read the card's input section, then prepare one identity vector of 128 numbers and one pose of 77 joint rotations for the SOMA-shape backend.

The core has no learned parts. Mesh transfer is a sparse matrix product, the skeleton comes from radial basis function regression with Kabsch alignment, and skinning is standard LBS with joint orient poses. The shape space is 128 principal components fitted offline from body scans. The optional corrective network has two hidden layers and about 10^8 parameters when enabled. The interface works with torch.export.

Identity coefficients for one backend
Barycentric transfer to the shared mesh
RBF fit of the 77 joint skeleton
Linear blend skinning with your pose
Posed body mesh comes out

Good to know: On an A100 80GB at batch size 128, the card reports more than 7,033 posed meshes per second. Single mesh times are below.

Title: Time per mesh in milliseconds at batch size 1
GPU: 2.1 ms
CPU, 32 cores: 12.1 ms

• License: Apache 2.0, and the card says the model is ready for commercial use
• Cost to run: none named
• Data it reaches: takes numeric tensors only, and MANO hand models are ones you supply yourself
• Not for: identity values far outside a backend's shape space, which can give odd geometry

On Hugging Face, 15 November 2026

• Task: robotics
• Downloads in the last 30 days: 0
• Likes: 38
• License: apache-2.0
• Tags: soma-x, parametric-human-body-model, computer-graphics, human-pose-estimation, animation, robotics, pytorch

Links

• Model: https://huggingface.co/nvidia/SOMA-X
• Files: https://huggingface.co/nvidia/SOMA-X/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download nvidia/SOMA-X

Then follow the usage in its model card. Nothing in this kit runs by itself.