Load NLF Model

This is the loader for NLF - Neural Localizer Fields, a monocular 3D human pose and body estimator. If you're doing pose-driven Wan work (UniAnimate, motion transfer, anything where a driving video's poses steer the generated character), you first have to get those poses out of the driving footage. NLF is one of the pose extractors this pack supports, and it works in full 3D rather than the flat 2D skeletons of classic OpenPose/DWPose. LoadNLFModel puts that model into memory so the pose-prediction and drawing nodes downstream can use it.

The appeal of a 3D pose model is exactly what the skeleton-based approaches lose: depth and body orientation. A flat 2D skeleton can't tell forward-lean from backward-lean; a 3D estimator can. That richer pose is useful when you want motion transfer that doesn't distort as badly on turns and depth changes.

How it works

It loads the NLF model weights and returns an NLFMODEL handle. That handle then feeds the prediction step (which runs the model on your driving images to get 3D pose predictions) and, ultimately, a draw node that renders those predictions into pose images your Wan pose-control path can consume. This node is purely the load - it doesn't process any images itself.

The inputs and output that matter

• nlf_model (enum, required) - which NLF model file to load, picked from what's available in your models folder. If the dropdown is empty, you haven't downloaded the model yet (see install).
• warmup (BOOLEAN, optional, default true) - runs a warmup pass on load so the first real inference isn't paying JIT/compile startup. Leave it on for smoother first-run timing; the only reason to turn it off is to shave load time when you're just wiring things up.

The output is nlf_model (NLFMODEL) - feed it into the NLF pose-prediction node.

How to install it

ComfyUI Manager: search ComfyUI-WanVideoWrapper, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-WanVideoWrapper
pip install -r ComfyUI-WanVideoWrapper/requirements.txt

then restart. You need the NLF model weights downloaded into the folder the loader scans - without them the nlf_model dropdown is empty and there's nothing to load. Make sure your WanVideoWrapper install is current, since NLF support is a relatively recent addition.

Common issues & troubleshooting

The model dropdown is empty. This is the #1 snag: an empty enum means the loader found no NLF model files. Download the weights and drop them where the node expects, then refresh/restart ComfyUI so it re-scans.

First inference is slow. That's the JIT/warmup cost. Keep warmup on (the default) so the model is compiled at load rather than stalling on your first real run.

It's a loader, not a preprocessor. LoadNLFModel by itself produces no poses - it only hands a model to the prediction node. If you're expecting pose images straight out of this, you're looking one node too early; the chain is load → predict → draw.