Kimodo

From the project README

Modly extension for NVIDIA Kimodo SOMA RP human motion generation.

This repository provides the extension shell that lets Modly prepare a Kimodo runtime, download the Kimodo SOMA RP weights through Modly's normal model-download flow, and expose a text-to-motion node inside the current mesh-centric Modly UI.

This is an initial private base, not a finished production integration.

What is wired today: Install / Repair prepares the extension runtime by running setup.py and creating the extension-local Python environment. Download uses Modly's existing Hugging Face download flow for nvidia/Kimodo-SOMA-RP-v1 and stores weights under Modly's configured models directory. The runtime exposes a Text to Motion Preview node with text -> mesh compatibility metadata. The runtime exposes an Animate Rigged Mesh model node with text + rigged mesh -> mesh compatibility metadata. It shares the same Kimodo weight ownership as the preview node. Kimodo's real output remains motion-first. For Text to Motion Preview, the .glb returned to Modly is still a preview/compatibility artifact and the canonical output is a motion bundle on disk. For Animate Rigged Mesh, the supported MVP path now attempts a real BVH-first retarget/export and returns animated.glb only after animation data is structurally authored into the GLB.

Requirements Modly with extension support. A Hugging Face token configured in Modly. Kimodo depends on gated Hugging Face resources. Kimodo's text encoder uses LLM2Vec on top of Meta Llama 3: McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp, McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp-supervised, and meta-llama/Meta-Llama-3-8B-Instruct. Before first use, sign in to Hugging Face with the token used by Modly and accept or request the required Meta Llama 3 access conditions before download; otherwise encoder/model download can fail even when Kimodo itself is installed. NVIDIA GPU runtime. Linux x8664 is the intended supported baseline. Linux arm64 + NVIDIA Blackwell/CUDA is currently an experimental host lane, validated locally on NVIDIA GB10 with pinned PyTorch ARM64 CUDA wheels. Windows x8664 is available as an experimental compatibility lane. The setup path is unit-tested for Windows-style virtualenv paths, but native Windows CUDA/PyTorch/DLL execution has not yet been validated by this repository.