Nodes/Link Comfy Nodes/HybridMaMoMask Loader
ComfyUI Node

HybridMaMoMask Loader

Text-to-3D-motion in ComfyUI starts here — and the first run downloads about a gigabyte

By Mister-Link·Created 10 months ago·Updated 6 days ago· 0
HybridMaMoMask Loader
    • model
    offload_to_cpufalse

    This is the node that turns ComfyUI into a text-to-motion tool: type a sentence, get a 3D human skeleton doing it. Not video generation - actual motion data, a 22-joint skeleton with per-frame rotations you can hand to a rig. The HybridMaMoMask loader is the boring first step of that chain, and it does exactly two things: pull the model files down and load them into memory. Every other HybridMaMoMask node depends on its model output, so you're not going anywhere without it.

    What you're actually loading

    The model behind it is HybridMaMoMask, a "Hybrid Transformer-Mamba for Text-to-3D Human Motion Generation" - an M.Tech thesis project (Chetana Dubey, IIT Gandhinagar) released as the Hugging Face Space Chetanaa/T2M under MIT. The README in this pack is a single line ("AI slop nodes for my own purposes, use at your own risk"), so the real docs live on that Space. The pipeline is four models: a length estimator, the hybrid Mamba+Transformer mask transformer, a residual transformer, and an RVQ-VAE decoder.

    On first use the loader runs snapshot_download from the Space into ComfyUI/models/hybrid_mamomask/ - that's the gigabyte of checkpoints (net_best_fid.tar files, VQ metadata, all of it). It only downloads once, but plan for it to take a while and to fail if you can't reach Hugging Face.

    Inputs and outputs

    One input, one output, that's the whole node:

    • offload_to_cpu (boolean, default false) - keep the model on the GPU for speed, or force CPU if VRAM is tight. It's four small-ish checkpoints, so most people leave this off.

    Output is a single model of type HYBRID_MAMOMASK_MODEL, which feeds straight into HybridMaMoMask Generate. That's its only job.

    Installing it

    It ships in Link Comfy Nodes, so:

    cd ComfyUI/custom_nodes
    git clone https://github.com/Mister-Link/link-comfy-nodes
    cd link-comfy-nodes
    pip install -r requirements.txt
    

    Or use ComfyUI Manager and search "Link Comfy Nodes", then restart. Note that requirements.txt is heavy - torch, transformers, opencv, scikit-learn, numba, and critically clip (installed from OpenAI's git) plus mambapy, einops, ftfy and regex. Those last four are the ones this node's runtime check insists on, and they're not things ComfyUI core installs.

    Common issues

    • "HybridMaMoMask is missing Python dependencies: clip, mambapy, einops..." - the node checks its imports and tells you exactly what's missing. Run pip install -r requirements.txt again; you likely skipped it or it failed partway.
    • First load is slow / looks hung. It's downloading the model repo. Watch your network, not the node.
    • "AI slop nodes... use at your own risk" isn't a joke. The author warns behavior changes without warning - pin your install or at least don't build a fragile production pipeline on the exact node defaults.

    Once the loader's done its big first download, the fun part is the Generate node.

    CategoryMotion/T2M

    Inputs (1)

    NameTypeDefaultDescription
    offload_to_cpuBOOLEANfalse

    Outputs (1)

    NameTypeDescription
    modelHYBRID_MAMOMASK_MODEL