What shipped
MotionBricks is an inbetweening model. Give it the four most recent poses of a character and up to four poses the character should reach, and it generates the motion between them at 30 frames per second. One generation covers 6 to 16 tokens of four poses each, which is 24 to 64 raw frames; after the boundary frames are removed, 18 to 58 valid frames remain, roughly 0.6 to 1.9 seconds. Longer sequences are built by chaining: the runtime plays what was generated, takes the last four poses, and uses them as the context for the next call.
The repository ships two pretrained models on the same runtime. MotionBricks-SOMA-RP-v1 animates a 30-joint human skeleton. MotionBricks-G1-RP-v1 generates kinematic poses for the Unitree G1 humanoid robot. Around them sit two demos, an Inbetweener and a Scene Preview, and two editors for authoring Smart Primitives, the target-pose recipes that turn one inbetweening model into vaulting, sitting and other interactions without retraining.
The SOMA checkpoint is about 200 million parameters in total. The pose transformer accounts for 136.5 million of them, the root model, which predicts duration and path, for 40.4 million, and the pose VQ-VAE for 23.4 million. The components fit the job: this is a component for an animation or simulation pipeline, not a foundation model asking for a data center.
Where the 15,000 frames per second claim lands
Secondary coverage of the paper led with speed: 15,000 frames per second, and about 2 milliseconds of latency. Those are different measurements, throughput and latency, and neither has been verified for the shipped ONNX files. The release documents something narrower and more useful. Timings are reported only for a GeForce RTX 4090 under Windows, through TensorRT 10.8 plans at several batch sizes, and the model card is explicit that a CPU comparison is not provided and that the tested setup is not a complete compatibility matrix.
That distinction matters for anyone budgeting an integration. A throughput figure from a paper tells you the method is fast in principle. A timing table tied to one named GPU, one operating system and one runtime version tells you what to expect on your desk, and what you still have to measure yourself on anything else. MotionBricks now offers the second kind of evidence. It should be read as the second kind, too.
The missing number
The SOMA model card is unusually direct about what it does not contain. A complete end-to-end benchmark of these ONNX models is not provided. Quantitative motion quality is not reported. An FP32-versus-mixed-FP16 quality comparison is not reported, even though the card notes that the two precisions can produce different motion because rounding changes which motion tokens get chosen, and that differences between the two precisions can accumulate when generated motion is fed back as context.
Training did measure some things: tokenizer reconstruction, foot skating, root-predictor validation and pose-predictor training loss. The card states that those measurements do not evaluate the exported ONNX models. The paper's experiments, it adds, describe the method on its own models and datasets, not a benchmark of these files.
So the honest position for an animator or robotics team is this: you can run MotionBricks today, on your own poses, in your own scene, and that direct test is currently the only published quality check available for the shipped SOMA model, because NVIDIA has published none. NVIDIA has, to its credit, written that down instead of leaving a benchmark-shaped gap for readers to fill with the paper's numbers.
The limits are part of the product
The limitations section deserves reading before the demos. Conditioning comes from poses and motion constraints only. Text prompts and scene geometry are not inputs; your application places the targets and handles the obstacles. The model works best with target poses inside its training distribution, and poses outside it can produce lower-quality motion. The SOMA layout is fixed at 30 joints, so other skeletons need a retargeting step.
For the robot model, the card is blunt: the model generates kinematic poses, and dynamic balance, actuator limits, collision avoidance and contact tracking require a controller downstream; generated motion is not validated for physical robot control. A G1 animated by MotionBricks still needs that controller and its own validation before anything physical follows from a generated pose sequence.
Who should try it, and on what terms
Treat this as a prototyping component with checkable speed and unmeasured quality. If you animate characters or build humanoid simulation, the cost of finding out is a clone and a test run: run the Inbetweener on poses from your own project, and judge the joins yourself, especially foot contact and what happens after several chained generations, where the card says precision differences can accumulate. If you need a quality guarantee for a production pipeline, or a controller for a physical robot, this release gives you neither, and says so.
Sources
- [1] NVlabs/motionbricks repository (README, setup, checkpoints, licenses)Read source
- [2] MotionBricks SOMA model card (parameters, training data, evaluation limits, limitations, timings scope)Read source
- [3] MotionBricks Unitree G1 model card (kinematic-pose scope, controller requirements, not validated for physical robot control)Read source
- [4] Secondary coverage of the standalone release, used for release context onlyRead source
- [5] Secondary release listing (repository publication dated November 19, 2026)Read source