Kimodo Load Model

You've probably used ComfyUI to make images, maybe video. This node is the odd new neighbor: it loads a model that makes motion - 3D joint positions and rotations for a human (or a Unitree robot) - from a text description. Kimodo_LoadModel is the first step of every workflow in the GuardSkill/ComfyUI-Kimodo-Bridge pack, which wraps NVIDIA's open-source Kimodo text-to-motion diffusion model.

Where it sits in the chain: Load Model → Text Encode → Sampler → Post Process → Preview/Export. Nothing else works until this node has a model in VRAM, so treat it as your gate. The first run will also be the slowest - it has to download a lot.

What it does

The node has exactly one input, model, a dropdown of five variants:

• Kimodo-SOMA-RP-v1 (default) - human, 30 joints, trained on the 700h Rigplay dataset. This is the one you want for everyday character work.
• Kimodo-SMPLX-RP-v1 - SMPLX human, for when you need that skeleton's joint layout.
• Kimodo-G1-RP-v1 - Unitree G1 humanoid robot, 34 joints.
• Kimodo-SOMA-SEED-v1 / Kimodo-G1-SEED-v1 - the same two skeletons trained on the smaller 288h SEED dataset.

Output is a single model (type KIMODO_MODEL) that you wire into Kimodo Text Encode. If you're doing a human character, leave it on SOMA-RP-v1 and move on; you'd mostly switch skeletons when a downstream exporter demands a specific layout.

The mechanism and the honest cost

"Auto-downloads from HuggingFace on first use" is doing a lot of work in that tooltip. The diffusion weights aren't the whole story - Kimodo's text encoder is built on Meta Llama 3 8B plus LLM2Vec adapters, and that encoder is a big chunk of why Kimodo wants roughly 17GB VRAM. Community reaction when Kimodo landed was mostly people doing the math on that number and sighing; it runs fine on 24GB cards, and this pack is really aimed at people with that kind of headroom. Models cache under models/Kimodo (the pack registers that folder with ComfyUI), so the download only bites once.

One gotcha the README is upfront about: Llama 3 is a gated model on HuggingFace. If the text-encoder download fails or stalls, you need to request access at the Meta Llama 3 page, create an HF token, and run huggingface-cli login before retrying. The diffusion model itself (nvidia/Kimodo-...) doesn't have that gate.

Installing the pack

The nodes install together with the whole pack. Easiest: ComfyUI Manager, search Kimodo Motion Bridge, install, restart. By hand:

cd ComfyUI/custom_nodes
git clone https://github.com/GuardSkill/ComfyUI-Kimodo-Bridge.git
cd ComfyUI-Kimodo-Bridge
python -m pip install -r requirements.txt
python install.py

Then restart ComfyUI and look under the Kimodo category.

Where people get burned

• First run looks hung. The model (plural - text encoder + diffusion weights) is downloading and then loading; watch the console for the [Kimodo] progress lines rather than judging by the UI.
• Llama access denied. Request the gate, log in with huggingface-cli login, re-queue.
• VRAM. If you're under ~20GB of VRAM, generation will likely OOM or crawl. The heavy encoder is the reason, not the small diffusion core - and no, that's not really your fault.

Load model first, confirm the [Kimodo] Model loaded: ... (skeleton=..., fps=...) line in the console, then build the rest of the graph on top of it.