ARDY-SOMA-RP-30FPS-KimodoRep-Horizon60: File List

ARDY-SOMA-RP-30FPS-KimodoRep-Horizon60: File List

At a Glance

Item Value
Repository nvidia/ARDY-SOMA-RP-30FPS-KimodoRep-Horizon60
Publisher guide NVIDIA: models and licenses
Published 2026-10-14
License nvidia-open-model-agreement
Formats safetensors
Source type Primary source (the publisher itself)

Values determined by this site’s code when the information was collected. Dates are JST.

Overview

NVIDIA has released a new variant of “ARDY," an autoregressive diffusion model designed for interactive human motion generation. This model accepts online text prompts and flexible long-term kinematic constraints (such as root paths, waypoints, full-body keyframes, or sparse joint positions and rotations) as inputs, generating pose sequences in real time.

The newly released ARDY-SOMA-RP-30FPS-KimodoRep-Horizon60 is a model developed as part of the ARDY project. It is trained at 30 fps using the Bones Rigplay 1 dataset with a 30-joint SOMA body model skeleton. Unlike other ARDY models that use raw hip trajectories as roots, this variant is designed to predict the smooth root motion representations used in the Kimodo model. This enables direct comparison with the root constraints and outputs of the Kimodo model.

Specifications

Model specifications are as follows:

  • Architecture: Diffusion model (Novel Two-Stage Transformer)
  • Parameter Count: 192 M (denoiser: 157 M, motion tokenizer: 36 M; text encoder parameters are not included)
  • Output Specifications: 30 fps, 30-joint SOMA body model skeleton, pose sequences up to 8 seconds in length (including global root movement and joint rotations)
  • Input Formats: Text, pose constraints, historical poses (maximum duration of historical poses is 8 seconds)
  • License: NVIDIA Open Model Agreement (available for commercial and non-commercial use)

Performance and Quality

The following items are listed as evaluation metrics for the model:

  • Pose Constraint Accuracy (pose constraint accuracy based on joint distance errors)
  • Motion Quality (motion quality based on foot-skating error, FID, and latent similarity)
  • Text-Following Accuracy (text-following accuracy based on R-precision and latent similarity)

Specific benchmark numerical scores are provided in the codebase within the repository.

Regarding reports on model performance, several technical limitations are also explicitly noted. Generated motions may contain artifacts such as “foot skating," where feet slide unnaturally when they should be in contact with the ground. Additionally, the generated motions do not necessarily follow the given text prompts completely, and the model may not have learned certain types of motions.

Note that this model is reported to perform exceptionally well in walking, gesturing, combat, dancing, and daily activities. On the other hand, it cannot generate cartoonish or physically implausible movements. Furthermore, each trained model is currently designed to output motion for a single character skeleton. In addition, the model lacks the ability to recognize objects in the surrounding scene.

Strengths and Use Cases

This model is suited for generating controllable humanoid motion in applications that require real-time performance. Specifically, the following use cases are anticipated:

  • Motion planning for humanoid robots
  • Character locomotion in digital twins and industrial simulations
  • Digital human motion generation for synthetic data creation
  • Animation in games and other interactive applications

As a characteristic of the model, it is noted for excelling in the generation of locomotion, gestures, combat, dancing, and daily activities. Input formats allow for control using text instructions (text-to-motion), as well as pose constraints and historical poses.

Distributed Files

Weight files published in nvidia/ARDY-SOMA-RP-30FPS-KimodoRep-Horizon60, listed by this site from the Hugging Face API. Sizes are the actual file sizes.

File Size
denoiser.safetensors 627MB
tokenizer.safetensors 142MB

How to Get It

The model is distributed in the following format:

PyTorch is used for operation. Additionally, acceleration using TensorRT is possible to speed up inference. For detailed installation and usage instructions, please refer to the official codebase.

Can You Run It Locally?

The publisher distributes this model as safetensors.

License — nvidia-open-model-agreement: A custom license from the publisher. Check the original terms directly, including whether commercial use is permitted.

Compiled by this site’s code from the published formats and the license field. License summaries are not legal advice — check the publisher’s original terms before relying on them.

Our Own Measurements

Measurements We Did Not Take

We have not confirmed that this site meets the commercial-use terms of this model’s license (nvidia-open-model-agreement). Because this site carries advertising, we did not run the model (no CPU run, answers, quantization comparison, conversion or generation).

Related Articles

What to Read Next

Sources