(Down)load CogVideo ControlNet

This node fetches a ControlNet for CogVideoX so you can steer the generation with structure - an edge map or a HED sketch - instead of relying on the text prompt alone. It's the loader half of a two-node pair: this one downloads and loads the ControlNet weights, and the CogVideoControlNet node then applies them with your control images. On its own this node just gets the model into memory.

If you've used ControlNet for images, the concept carries straight over: give the model a per-frame structural guide (here, canny edges or HED soft edges) and it tries to match that structure while filling in the content from your prompt. For video that's how you get motion that follows a specific shape or outline rather than whatever the model felt like doing.

What it loads

The model dropdown has four entries, all from TheDenk (the author of these CogVideoX ControlNets):

• cogvideox-2b-controlnet-canny-v1 and cogvideox-5b-controlnet-canny-v1 - canny edge control. Hard, precise outlines. Good when you want the model to stick tightly to a defined shape.
• cogvideox-2b-controlnet-hed-v1 and cogvideox-5b-controlnet-hed-v1 - HED soft-edge control. Looser, more organic contours. Better when hard canny lines feel too rigid.

Pick the one that matches your base model's size - a 2b ControlNet for a 2b CogVideoX model, 5b for 5b. That's the pairing that matters; mixing sizes is asking for trouble.

The one output, cogvideo_controlnet (a COGVIDECONTROLNETMODEL), goes into the CogVideoControlNet node, where you combine it with your control image sequence before it reaches the sampler.

How to install it

It ships with the wrapper, and it downloads the ControlNet weights itself on first run:

• ComfyUI Manager - search ComfyUI-CogVideoXWrapper, install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/kijai/ComfyUI-CogVideoXWrapper, then pip install -r ComfyUI-CogVideoXWrapper/requirements.txt, restart.

To make the control images themselves - the canny or HED maps - you'll usually want a preprocessor pack like comfyui_controlnet_aux to convert your source video into edge sequences.

Common issues & troubleshooting

Only canny and HED, that's it. These are the only control types available for CogVideoX in this wrapper - there's no depth, pose, or tile ControlNet here. If you need those, CogVideoX isn't the model for it; that's one of the areas where the newer video models and their VACE-style control have pulled ahead.

Loading it does nothing visible. Right - this node only loads the model. Nothing changes in your output until you feed its output into CogVideoControlNet with actual control images and route that into the sampler's controlnet slot. The loader is step one of two.

Mismatched model size. A 5b ControlNet on a 2b model (or vice versa) won't behave. Match the ControlNet's size to your CogVideoX model.