Wan Video Prompt Extender Select

Wan responds best to long, specific, cinematic prompts - the KB's rule of thumb is "for image you are a robot taking pictures, for i2v you are George R.R. Martin." Most people type three words and wonder why the motion is boring. This node wires in Wan's official prompt extender: a small LLM that takes your short prompt and rewrites it into the rich, detailed description the model actually wants. Alibaba ships these exact system prompts with Wan; this node just lets you pick one and feed it into the workflow.

It's a config selector, not the generation itself. You choose which rewriting behavior you want, and it outputs an extender_args bundle that the sampling/encode path uses to run the rewrite at generation time.

How it works

You pick an LLM model and one of six built-in system_prompt presets. Each preset is a different rewriting personality tuned for a Wan mode. At generation time the model reads your original prompt through that system prompt and expands it. The presets tell the story of what's on offer:

• T2V Movie Director (English / Chinese) - for text-to-video: turns a plain idea into a directed, shot-described scene.
• I2V Rewriter (English / Chinese) - for image-to-video: rewrites your prompt to describe motion applied to an existing image.
• I2V Imagination (English / Chinese) - a looser I2V mode that invents more, for when you want the model to run with the idea.

The English/Chinese split is because Wan was trained bilingually and both work; pick the language you prompt in.

The inputs and outputs that matter

• model - the LLM used to do the rewriting, from the models it finds on disk. Required.
• system_prompt - the preset (the six above). This is the real choice: match it to your mode. Using a T2V director prompt on an I2V job gives you a rewrite that describes a scene from scratch instead of describing motion on your image.
• max_new_tokens (default 512) - how long the rewrite can get. 512 is plenty; raise it only if your expansions are getting cut off.
• custom_system_prompt (optional STRING) - override the presets entirely with your own instructions. Leave empty to use the selected preset.
• seed (optional) - the LLM is sampled, so the seed makes a rewrite reproducible. Handy when you found an expansion you liked.

Output: extender_args (WANVIDEOPROMPTEXTENDER_ARGS) → the matching input on the sampler / text-encode node.

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. The catch: this loads an actual LLM for the rewriting (a Qwen-class model is the usual pick), which has to be downloaded and takes its own VRAM/RAM. On a tight card, running the LLM and Wan back to back can be the thing that tips you into swapping.

Common issues & troubleshooting

The rewrite goes off the rails. Wrong preset for your mode is the usual cause - an "Imagination" preset is meant to invent, so it'll add things you didn't ask for. Switch to the plainer "Rewriter" preset, or write a custom_system_prompt that pins down what to keep.

No change in output. Make sure extender_args is actually wired into the sampler and that the workflow runs the extender before encoding text. If it's dangling, your original short prompt is what gets used.

It's slow / you OOM. You're loading a second model. If you don't need automatic rewriting, skip the node and just write a long prompt by hand - the extender is a convenience, not a requirement, and a well-written manual prompt beats a mediocre auto-expansion.

Rewrites change every run. Set seed to lock them once you're happy.