SCAIL CMU Library Loader

The CMU Motion Capture Database is one of the oldest and biggest public mocap collections around - 2,500+ sequences covering everything from walking and running to fighting, climbing, swimming and two-person interactions. SCAILCMULibraryLoader is how this pack gets at it. It downloads, indexes and hands over that library so SCAILCMUMotion can play individual sequences, and it's the node that makes the non-dance half of this pack possible.

Two inputs:

• source - huggingface or local. Default behavior: if you don't touch it, it goes to HuggingFace.
• local_path - an optional string used when source is local, pointing at a folder containing the CMU .npy files plus an index.json. Leave it empty with huggingface selected and it auto-downloads.

The output is a single library of type CMU_LIBRARY, which is the required input for SCAILCMUMotion. That's the only consumer - this isn't a general-purpose dataset node, it's plumbing for one sibling.

The download you should plan for

On first use with huggingface selected, it pulls cmu_npy.tar.gz (~660MB) from the ckinpdx/CMUMDB dataset on HuggingFace into ComfyUI/models/cmu_motion, verifies the file size (it retries if the download looks truncated, up to three attempts), and extracts. The console shows the progress. Unlike the AIST loaders there's no toggle to disable the download - if the index isn't there, it downloads, full stop. Budget for it: this is the biggest single download in the pack, and a flaky connection will make it feel endless. If you have the bandwidth, grab it on purpose before you need it by just running the loader once and letting it finish.

The local option is the escape hatch for exactly that scenario: download the tarball elsewhere, extract it, point local_path at the folder with index.json, and the loader skips networking entirely.

Installing and licensing note

Pack install is shared: clone ckinpdx/ComfyUI-SCAIL-AudioReactive into ComfyUI/custom_nodes, pip install -r requirements.txt, restart, or use ComfyUI Manager. No extra Python deps beyond the shared taichi/librosa stack.

One thing the README is upfront about: the CMU data is free for use including commercial, but the database asks for a citation in published work (the specific line is in the pack's README, crediting NSF funding for the database). It's a two-line attribution, not a legal maze, but if you publish generated content built on this library, throw the citation in the description. The AIST++ data has a similar CC-BY 4.0 attribution requirement - this pack is great, but you're responsible for what its data does downstream.