All notable changes to this project will be documented in this file.
The format is based on Keep a Changelog.
- Option
--bvh_standard_tposeto use standard T-pose for BVH file saved fromgenerate.py - Option to use standard T-pose for BVH file saved or downloaded from demo
- Option to input/output BVH files that use standard T-pose with
motion_convert.py - Added BVH file containing the standard Kimodo T-pose to
kimodo/assets/skeletons/somaskel77/somaskel77_standard_tpose.bvh - Updated documentation with these new options
- Kimodo-SOMA-RP-v1.1 and Kimodo-SOMA-SEED-v1.1 models and added support in the codebase. If not specified, the latest version of the models will be used automatically with the demo and CLI.
- Kimodo Motion Generation Benchmark for standardized evaluation of motion generation models training on the BONES-SEED dataset.
- Scripts to construct the full benchmark, generate motions for test cases, and compute evaluation metrics.
- Documentation explaining the benchmark and how to use the evaluation pipeline.
- TMR-SOMA-RP-v1 motion-text embedding model to be used for evaluation metrics.
- Added option to load LLM2Vec text encoder in fp32 precision.
- Always use batch size 1 with LLM2Vec to avoid unexpected behavior of different embeddings based on batch size.
- Load LLM2Vec directly onto the GPU, if available.
- Updated documentation on constraints with more details.
- Fix unnecessary text encoder reload when switching between models in the interactive demo (if not using the text encoder server API).
- New
kimodo_convertCLI tool for converting generated motions between formats (NPZ, BVH, MuJoCo CSV, AMASS NPZ). - Support for loading and saving BVH, CSV, and NPZ motion files in the interactive demo.
- Bug fix for foot contact visualization in the interactive demo.
- Patch bug with BVH export for SOMA models.
- Breaking: Model inputs/outputs now use the SOMA 77-joint skeleton (
somaskel77). This affects saved motion formats and constraint files from previous versions.
- Released timeline annotations for the BONES-SEED dataset on HuggingFace.
- Open-source release of Kimodo codebase under Apache-2.0 license.
- Five model variants: Kimodo-SOMA-RP-v1, Kimodo-G1-RP-v1, Kimodo-SOMA-SEED-v1, Kimodo-G1-SEED-v1, Kimodo-SMPLX-RP-v1.
- Command-line interface (
kimodo_gen) for motion generation with text prompts and kinematic constraints. - Interactive web-based motion authoring demo (
kimodo_demo) with timeline editor, constraint tracks, and 3D visualization. - Support for multiple output formats: default NPZ, MuJoCo qpos CSV (G1), AMASS NPZ (SMPL-X).
- Documentation site with quick start guide, installation instructions, CLI reference, and API docs.
- Compatibility with downstream tools: ProtoMotions (physics-based policy training) and GMR (motion retargeting).