Qwen Loader

Quick clarification up front, because the name trips people up: this is a loader for the Qwen large language model - the text LLM - not Qwen-Image or any diffusion model. In the WanVideoWrapper it exists for one purpose: to feed the WanVideoPromptExtender node, which uses an LLM to rewrite your short prompt into the long, descriptive style Wan actually likes. QwenLoader gets that LLM into memory and hands it downstream as a QWENMODEL. On its own it does nothing visible; it's the engine you bolt onto the prompt extender.

Wan was trained on verbose, cinematic captions, and short prompts genuinely underperform on it. An LLM that expands "a woman walking on a beach" into a paragraph of camera, lighting, and motion detail is a real quality lever - and this is the node that supplies that LLM.

How it works

It loads the Qwen model weights at the precision and device you pick and exposes the loaded model. That's the whole job - no generation happens here. The actual prompt rewriting is done by WanVideoPromptExtender, which takes this QWENMODEL plus your prompt and runs the LLM to produce expanded text.

The inputs and outputs that matter

• model (enum) - which Qwen model to load, populated from your local models folder. If the dropdown is empty, you haven't downloaded a Qwen model yet - the classic first-run blocker.
• precision (bf16 default, or fp16 / fp32) - load precision. bf16 is a sensible default for an LLM; drop to fp16 if your setup prefers it.
• load_device (main_device / offload_device) - where the model sits. LLMs are chunky, so if you're running it alongside a 14B Wan model on a tight card, offload_device keeps VRAM free for the diffusion work.

Output is QWENMODEL - wire it straight into WanVideoPromptExtender.

How to install it

Comes with the WanVideoWrapper. 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 need a Qwen LLM downloaded into your models folder for the dropdown to populate - it's a separate, sizable download (an LLM, not a tiny adapter). Budget disk and RAM/VRAM accordingly.

Common issues & troubleshooting

Empty dropdown. No Qwen model found. Download one into the right folder and refresh. Same first-run story as every loader in this pack.

VRAM blows up when you also load Wan. An LLM plus a 14B video model is a lot to hold at once. Set load_device to offload_device here, and note the prompt extender can offload after it runs - you don't need Qwen resident while the actual video samples.

Loaded but nothing happens. Expected. QwenLoader produces no text by itself. It's inert until WanVideoPromptExtender uses it. If you wanted prompt expansion, wire the extender to this node's output.