The two-stage transformer, motion representation and scheduler are adapted from:
Kimodo, Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES, Apache-2.0.
https://github.com/nv-tlabs/kimodo
Reference revision: 58e781898b3d7e328a676a75d3e338c45dce3ad9
Text tokenization and pooling follow LLM2Vec, Copyright (c) 2024 McGill NLP, MIT.
The package contains no learned weights. Converted checkpoint licensing is separate.