WanVideo T5 Text Encoder Loader

Wan reads your prompt through a T5 text encoder - specifically the umT5-XXL model - and this node is how you load it inside Kijai's wrapper. It's a small, boring, load-bearing node: without it, WanVideoTextEncode has nothing to encode with, and your prompt never reaches the model. You wire the output straight into the text-encode node and mostly forget it exists.

The node's own description tells you where it looks: ComfyUI/models/LLM. That's the folder to drop the encoder file into. This is a separate download from the diffusion model itself - a common trip-up for people who grabbed the checkpoint and expected the whole thing to just run.

How it works

The text encoder turns your words into the conditioning embeddings that steer denoising. Wan uses umT5-XXL (a multilingual T5), which is a chunky model in its own right - several gigabytes - so precision matters for VRAM. It's loaded once and reused for every generation in the session.

The inputs and outputs that matter

• model_name - the encoder file, picked from models/LLM. Grab the umT5-XXL file Kijai packages alongside the Wan weights.
• precision - bf16 by default, which is the quality baseline. fp32 is heavier for no real benefit at inference time.
• quantization (optional) - disabled or fp8_e4m3fn. This is the one knob worth touching: loading the T5 in fp8 roughly halves its memory footprint with negligible quality difference for typical prompts. On a tight card, it's usually the first VRAM saving people make, and per the wider community consensus on fp8, the quality cost is close to invisible.
• load_device (optional) - offload_device (the default) keeps the encoder in system RAM until it's needed, freeing the GPU for the diffusion model. Leave it there unless you have VRAM to spare.

Output is a single WANTEXTENCODER, which feeds the t5 input of WanVideoTextEncode.

How to install it

Comes with the whole pack. Via ComfyUI Manager: search WanVideo Wrapper, install, restart. Manually:

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

then restart. The node is trivial; the umT5-XXL encoder file is the real download, and it goes in models/LLM, not models/text_encoders or models/clip.

Common issues & troubleshooting

The dropdown is empty or the file won't load. The encoder isn't in ComfyUI/models/LLM, or it's in the wrong folder. This node specifically reads from LLM - if you dropped the umT5 file into text_encoders (where native ComfyUI Wan nodes expect it), the wrapper won't see it. Move it and restart.

VRAM pressure with the model loaded too. Set quantization to fp8_e4m3fn and keep load_device on offload_device. Between an fp8 encoder and an fp8 diffusion model, you claw back a real chunk of memory.

Non-English prompts behaving oddly. umT5 is multilingual, so it does handle other languages, but Wan's training skews heavily English - expect best prompt adherence in English regardless of what the encoder technically supports.