NVIDIA Releases Kimodo-SOMA-RP-v1.2 for 3D Motion Generation

NVIDIA Releases Kimodo-SOMA-RP-v1.2 for 3D Motion Generation

At a Glance

Item Value
Repository nvidia/Kimodo-SOMA-RP-ArdyRep-v1.2
Publisher guide NVIDIA: models and licenses
Published 2026-10-14
License nvidia-open-model-license
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 “Kimodo-SOMA-RP-v1.2", a diffusion model that generates 3D skeleton animations based on text prompts and detailed motion constraints. This model can generate 3D skeleton motions tailored to inputs such as text instructions, full-body poses, joint end-effector positions, movement trajectories, and waypoints.

Specifications

  • Architecture: Diffusion Model (Novel Two-Stage Transformer)
  • Parameters: 282 M
  • Output Specifications: 30 fps, up to 10 seconds (300 frames)
  • Output Format Details:
    • Root Translation: 2D matrix (num_frames x 3) – Joint Rotations: 4D matrix (num_frames x 30 x 3 x 3)
  • Input Format Details:
    • Text: 1D string – Duration: 1D integer – Pose Constraints: 1D frame index for each constraint – Constrainable features include 3D joint positions, 3×3 joint rotation matrices, 2D heading directions, 2D root positions, and more.
  • Supported Skeleton: 30-joint SOMA skeleton
  • Training Data:
    • Bones Rigplay dataset (700 hours of motion capture data, text under 1 billion tokens)
  • License: NVIDIA Open Model License (reported to allow commercial use)

Performance and Quality

This v1.2 release fixes training splits. In the previous v1.1, some test data sequences leaked into the training set, meaning numbers on the Motion Generation Benchmark could not be fully compared with models trained on recommended splits. This leak has been removed in v1.2, making it a proper model for comparisons with other models on benchmarks.

Regarding performance changes compared to v1.1, the model card reports the following. In v1.2, unnatural foot sliding (“foot skating") is reduced, and constraint accuracy is maintained. Meanwhile, metrics indicating text adherence (R-precision and FID) are reported to be slightly lower than in v1.1, as a result of reflecting the removal of the data leak.

Strengths and Use Cases

This model aims to generate 3D human motion data for applications regardless of the user’s animation experience level. Specifically, the following use cases are anticipated:

  • Humanoid robot demonstrations
  • Digital human motions in digital twins and industrial simulations
  • Digital human motion for synthetic data generation
  • Animation creation in game and media development

The model excels at locomotion, gestures, combat, dancing, and everyday actions. On the other hand, it is not designed to generate cartoon-like movements or physically unnatural motions, nor can it recognize surrounding objects to adjust its movements.

How It Differs from Similar Models

As a variation of this release, nvidia/Kimodo-SOMA-RP-ArdyRep-v1.2 has been published. This model is designed to predict the Ardy motion representation, outputting global roots and joint rotations using raw hip trajectories instead of the smoothed roots used in other Kimodo models. This enables direct interoperability with the ARDY model.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch) — 283M parameters

Your VRAM Quantization File size Est. memory needed
4GB (laptop iGPU / phone class) F32 1.1GB 1.3GB

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. The listed files are distributed in F32 (32-bit). Loading in a lower precision (BF16 or quantized) needs less memory, so treat this as the size of the files as distributed, not a minimum. This site has not measured the lower-precision figures. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

Can You Run It Locally?

The publisher distributes this model as safetensors.

License — nvidia-open-model-license: 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-license). Because this site carries advertising, we did not run the model (no CPU run, answers, quantization comparison, conversion or generation).

Distributed Files

Weight files published in nvidia/Kimodo-SOMA-RP-v1.2, listed by this site from the Hugging Face API. Sizes are the actual file sizes.

File Size
model.safetensors 1.13GB

How to Get It

Related Articles

What to Read Next

Sources