WanVideo Controlnet Loader

Worth saying up front because it's genuinely a little unusual: this loads a real ControlNet architecture for Wan, not the thing most of the ecosystem uses instead. By the time Wan matured, "video ControlNet" mostly stopped meaning ControlNet at all - Wan's dominant control path runs through VACE, and before that the community leaned on spacepxl's control LoRAs trained on input/output video pairs. A proper, separately-trained ControlNet model for Wan2.1, the kind with its own weights and its own conditioning architecture, is the less common route. That's exactly what this node loads - models from TheDenk's Wan2.1 ControlNet collection on HuggingFace.

What it's for

If you've used ControlNet on image models - Canny, depth, pose, the usual conditioning types - this is the same basic idea ported to Wan by a third-party trainer rather than baked into an official Alibaba release. You load one of these purpose-trained models here, then feed the resulting WANVIDEOCONTROLNET object into whatever downstream node in your graph consumes it alongside your control video or image.

Worth knowing before you reach for it: since VACE is the path most of the Wan ecosystem actually uses for conditioned generation, and it's what most workflows, tutorials, and LoRA compatibility notes assume, this loader is more of a specialist tool. Reach for it specifically if you need a control type or a training recipe TheDenk's collection covers that you can't get through VACE, not as your default control mechanism.

The inputs and output that matter

model - which specific ControlNet checkpoint to load, from TheDenk's HuggingFace collection. base_precision (default bf16, also fp32/fp16) sets the compute precision. quantization (default disabled, also several fp8 variants including a "fast" and a "fast, no FFN" option) trades memory for a small quality cost - the fp8 options are worth trying first if you're VRAM-constrained, since Wan-scale models are large enough that quantization is often the difference between fitting on your card or not. load_device (default main_device, or offload_device) controls whether the model lives on GPU or gets offloaded to save VRAM at the cost of some speed.

Output is a single controlnet (WANVIDEOCONTROLNET) - the loaded model, ready to plug into whatever node in your graph applies the control conditioning.

How to install it

Via ComfyUI Manager, search WanVideoWrapper and install, or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-WanVideoWrapper

then restart. The actual ControlNet checkpoints aren't bundled - download them from TheDenk's wan21-controlnets collection on HuggingFace and place them wherever this loader's model dropdown expects them (typically your standard Wan model directory structure; check the dropdown after installing to confirm the exact path it scans).

Common issues & troubleshooting

Empty dropdown, nothing to select. If model shows no options, you haven't downloaded any of TheDenk's checkpoints into the folder this node scans yet - this is a loader, not a downloader, so the files have to already be on disk.

You expected VACE-level community support and aren't finding it. Because this is a third-party model outside Alibaba's official tooling and outside the VACE path most workflows are built around, expect thinner documentation and fewer example workflows than you'd get with VACE or Wan's native control options. Budget extra time to figure out conditioning types and strength by experiment.

Out-of-memory on a large Wan checkpoint plus a ControlNet loaded alongside it. Try fp8_e4m3fn quantization first, and if you're still tight, set load_device to offload_device - it costs some speed but keeps the ControlNet off GPU memory until it's actually needed.