kimodo.cpp Released: C++ Port Without PyTorch

The open-source project kimodo.cpp has been released, reimplementing NVIDIA's text-based 3D motion generation model Kimodo using only C++ and GGML, without PyTorch or the Python runtime. This implementation accepts UTF-8 prompts via a single executable file and generates joint rotations and root translations for the SMPL-X 22 skeleton, providing contrast tests that guarantee identical results on both CPU and Vulkan GPU.

By linking directly only to GGML and GGUF rather than the entire llama.cpp, the project directly implements the Bi-directional attention and average pooling operations required for the text encoder. This enables accurate inference without conflicting with existing Causal-style runtimes. Additionally, a memory optimization strategy was applied by sequentially loading the text encoder and denoiser, resulting in lower GPU memory usage compared to the original Python implementation.

Currently, the Kimodo-SMPLX-RP-v1 checkpoint is supported, while support for constraints, SOMA checkpoints, GLB export, and quantized models is scheduled for future updates. The project includes a rigorous contrast test suite based on fixed test data and metadata to verify numerical consistency with the original.