PartLock-CMU is a streaming text-to-motion model (a generator with 3050 million parameters + a part-wise motion tokenizer with 750 million parameters). When given short phrases like "walk," "jump," or "salsa dance," it continuously generates full-body motions. After switching sentences, the change takes effect in approximately 0.13 seconds, with smooth transitions between actions. The authors trained the model using only 2308 clips from the CMU motion capture dataset on a single RTX 3060 GPU; the entire training process took about 3 hours and 50 minutes. On the same GPU, each step (consisting of 4 frames) takes 35–38 milliseconds, which is approximately 3.5 times faster than real-time requirements. The output consists of positions and rotations for 22 joints at 30 fps (in SMPL order), requiring manual retargeting to character skeletons.
Why it matters Trained solely on the CMU motion capture dataset, the authors state it can be used in commercial products. It enables real-time generation on consumer-grade NVIDIA GPUs, and the output skeletal data can be retargeted to characters.
Hugging Face Likes 1 · Hugging Face Trending Score 1 · Hugging Face Downloads 0
License: Commercial use allowed · ComfyUI: Not available · VRAM: ≈1 GB (estimated)
Reasons
- Code and weights are licensed under Apache-2.0; the model card clearly documents training data, splits, and limitations, with uncurated examples.
- Includes a local real-time demo page; on Windows, it can be launched by double-clicking the script.
Usage
- After running
pip install -r requirements.txt, executepython live/live_server.pyand open http://localhost:8765 to input short phrases; or usepython generate.py --sec 4 --out motion.npz "walk" "jump"to export motion files.
Risks
- It only generates motions present in the CMU dataset; the authors found that nearly all of the 20 new actions they tested resulted in characters standing still. Mixing two actions in a single sentence yields poor results.
- Only supports English prompts; requires an NVIDIA GPU; output is skeletal data, not meshes; the demo interface is in Japanese.