Licensed by NVIDIA Corporation under the NVIDIA Open Model Agreement.
Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.

The agreement is included as LICENSE and published at
https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-agreement/

This repository repacks the SOMA checkpoints of nvidia/MotionBricks (soma/soma-v6-nova-4n8g-ea43dd8-fp32-v1) for
Diffusers. The float32 tensors of the root model, the pose model and the decoder of the pose VQ-VAE are unchanged,
renamed for the Diffusers model code next to them; the fused attention input projections were split into query, key
and value. The VQ-VAE encoder and the training-only codebook statistics were left out. The normalization statistics and
the neutral joint positions of the skeleton come from the same package and are stored as buffers, with the standard
deviations as sqrt(std^2 + 1e-5), as in NVIDIA's exported ONNX model.

The model code is new, Apache-2.0, and written by The HuggingFace Team.