HY-Motion Load Network

Everything else in the HY-Motion pack is text plumbing or file output. This is the model. Load Network is the node that pulls Tencent's HY-Motion 1.0 Motion Diffusion Transformer - the billion-parameter HunyuanMotionMMDiT - off disk and into VRAM, and its network output is what Generate actually runs. Think of the Load LLM nodes as the mouth and this as the brain.

You get one input and it's a dropdown: model_name, with two real choices. HY-Motion-1.0 is the full model - 27-layer MMDiT with 1280-dim features, ~8GB+ of VRAM for the weights alone, and the best instruction-following and motion quality. HY-Motion-1.0-Lite is the 18-layer/1024-dim version at ~4GB+, tuned down to run on consumer cards, and it's the default for a reason. On anything less than a 16GB card, Lite is the sane starting point. The README's own feature list frames the trade as "significant improvements in instruction-following and motion quality" for the big one - the kind of difference that matters once you're doing complex multi-action sentences, not for "a person is walking forward."

How it loads

The node scans ComfyUI/models/HY-Motion/ckpts/tencent/ for a folder containing both a config.yml and a latest.ckpt, then reads the config to build the network, loads the checkpoint (only the motion_transformer.* weights), and pulls the Mean/Std normalization stats from a stats/ subdirectory (or the checkpoint itself). It also pins a few generation constants you can't see but should know about: it works at 30fps, has a hard 360-frame window (that's the 12-second cap in Generate), and runs 50 diffusion/flow steps per sample. None of that is exposed as a knob, which is fine - you don't need it to be.

Where the models come from

Neither model ships with the pack. Download them from HuggingFace and place them like this:

ComfyUI/models/HY-Motion/ckpts/tencent/HY-Motion-1.0/{config.yml, latest.ckpt}
ComfyUI/models/HY-Motion/ckpts/tencent/HY-Motion-1.0-Lite/{config.yml, latest.ckpt}

or with huggingface-cli:

mkdir -p models/HY-Motion/ckpts/tencent
huggingface-cli download tencent/HY-Motion-1.0 --local-dir models/HY-Motion/ckpts/tencent

(If the folder lands flat instead of nested, fix the layout - the loader is looking for ckpts/tencent/<model_name>/.) The dropdown only shows what's actually on disk, so once it's downloaded and ComfyUI is restarted, your choice appears.

The one thing to remember

The network is only half the VRAM bill. You'll typically also have a Qwen3 encoder loaded (5-16GB depending on quantization) and, if you run the recommended prompt-rewrite path, the prompter model on top. A 12GB card running full HY-Motion-1.0 plus a full encoder is a recipe for an OOM that has nothing to do with this node. The combination people actually get working: HY-Motion-1.0-Lite + a quantized (GGUF or bnb-4bit/AWQ) encoder. If you hit memory errors, fix the encoder first - the diffusion model here is not the part that usually blows up.

Install the pack itself the usual way (ComfyUI Manager search "ComfyUI-HY-Motion1", or clone into custom_nodes/ and pip install -r requirements.txt), download a checkpoint, restart, and Load Network becomes the boring, reliable first node of every workflow.