Kimodo.cpp ports NVIDIA's text-to-motion model to GGML

Kimodo.cpp ported NVIDIA's Kimodo-SMPLX-RP-v1 text-to-motion model into GGML in 48 hours, according to its GitHub repo.

The port takes a model built for research GPUs and runs it from a plain text prompt to a moving SMPLX skeleton on ordinary CPU or Vulkan hardware, with parity tests confirming both backends produce matching output. Anyone who wanted this kind of motion generation before needed a CUDA GPU and a Python stack to drive it.

Under the hood it handles GGUF loading, safetensors conversion, and DDIM sampling, exposed through C and C++ APIs. It takes either a raw UTF-8 prompt or a precomputed LLM2Vec embedding, and the KIMODO_TEXT_LAYER_CHUNK environment variable tunes VRAM use, set between 1 and 32.

Constraints, SOMA support, G1 output, GLB export, and quantized models aren't built yet. That's most of what would turn a motion generator into a tool a studio could plug straight into its render workflow.

The repo passed 375 stars within two days of posting, fast enough to show real interest but too soon for anyone to have stress-tested the CPU path on a production-size scene.