Kimodo-SOMA-RP-v1.1

Kimodo-SOMA-RP-v1.1

Mirror of NVIDIA Kimodo-SOMA-RP-v1.1, a kinematic motion diffusion model that makes 3D skeletal animation from text and pose constraints.

What it is for

The repo holds the model weights (model.safetensors), its config.yaml, and the motion normalization statistics (mean and standard deviation for body, global root and local root). Kimodo generates skeletal body motion on the 30-joint SOMA skeleton from a text prompt and optional constraints: full-body poses, end-effector positions, paths and waypoints. The output is root translation and per-joint rotation matrices. This variant was trained on the proprietary Bones Rigplay dataset, with the test splits of the Kimodo Motion Generation Benchmark held out. Use it to make motion for humanoid robot demonstrations, digital humans in simulation and synthetic data, and game and media animation.

Provenance

• Upstream model: nvidia/Kimodo-SOMA-RP-v1.1. Code: nv-tlabs/kimodo.
• Paper: Kimodo technical report. Project page: Kimodo. Benchmark: Kimodo-Motion-Gen-Benchmark.

Use

hf download chibifire/Kimodo-SOMA-RP-v1.1 --local-dir Kimodo-SOMA-RP-v1.1

Load the weights with the kimodo library from the upstream code repo. See upstream's card for full usage, inputs, outputs, supported hardware and the bias, privacy and safety subcards.

Licence

The weights are under the NVIDIA Open Model License, the licence of nvidia/Kimodo-SOMA-RP-v1.1. LICENSE holds the agreement as upstream ships it, and NOTICE holds the attribution it asks for. Keep both with the weights when you redistribute them. The kimodo code repo is Apache 2.0; that licence covers the code, not these weights.