WanVideo CLIP Text Encoder Loader

Start with the trap, because the class name is genuinely misleading. It says "Text Encoder," but it does not load a text encoder. Its own description is the honest version: it "loads Wan clip_vision model from ComfyUI/models/clip_vision," and its only output is a CLIP_VISION. So this is the CLIP vision loader for Wan image-to-video - the thing that lets the model look at your reference image - not the T5 that handles your prompt. If you were hunting for prompt encoding, that's WanVideoTextEncodeCached. This node is for the image side of I2V.

Why you need it

Image-to-video is the community's default Wan workflow - anchoring the first frame to a known image solves half the consistency problems that plague pure text-to-video. But for the model to condition on that image, something has to encode it into a visual embedding, and that's the CLIP vision model. This loader is step one of that chain: load the CLIP vision weights here, then hand them to an encode node that actually looks at your picture.

How it works

It's a loader, so the mechanism is simple: point it at a CLIP vision file on disk, pick a precision, and it returns the loaded model. The one setting with real consequences is where it lives while idle.

The inputs and outputs that matter

• model_name - the CLIP vision file to load, read from ComfyUI/models/clip_vision.
• precision (fp16 default / fp32 / bf16) - fp16 is fine here; this model is small next to the diffusion weights, so there's little reason to reach for fp32.
• load_device (offload_device default / main_device) - offload keeps it in system RAM until it's needed, which is the right call on a tight GPU. Force it to main_device only if you've got VRAM to spare and want to skip the shuffle.

The single output, wan_clip_vision (CLIP_VISION), feeds straight into WanVideoClipVisionEncode (or the older WanVideoImageClipEncode), which is where your actual reference image goes in.

How to install it

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

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

then restart. You'll need the CLIP vision model file itself in ComfyUI/models/clip_vision - the Wan I2V CLIP vision weights from Kijai's WanVideo_comfy HuggingFace repo - before the dropdown shows anything.

Common issues & troubleshooting

Empty model_name dropdown. The node scans ComfyUI/models/clip_vision specifically. If your CLIP vision file is sitting in models/clip or somewhere else, it won't appear - move it into clip_vision and restart.

"But it says Text Encoder." It doesn't matter what the class name implies; trust the output type. CLIP_VISION out means image conditioning, full stop. Don't wire it anywhere a text embedding is expected.

You're on a text-to-video graph and wondering why you'd need this. You don't. CLIP vision is only in play for I2V (and image-conditioned variants). A pure T2V run skips this node entirely and gets its guidance from the text encoder alone.