FunPack Prompt Enhancer (Standalone)

FunPack Prompt Enhancer is the standalone version of the LLM that sits inside FunPack Studio. Feed it a rough idea - "a dragon over a city at dusk" - and it runs that through an actual language model to produce a rich, structured video description, ready for your sampler. It's the node to reach for when you know what you want but the two-line prompt you'd type isn't giving the model enough to work with.

How it works

The mechanics are straightforward: this is a local LLM, not an API call. You give it a user_prompt and a system_prompt (the default one is a carefully-worded instruction to describe the video's content, color, action, background, lighting, and camera moves - exactly the axes video models care about). The model rewrites your idea, and the result comes out the enhanced_prompt output.

Two model sources, picked by model_path_type:

• HuggingFace Pretrained - the default. It pulls from a HuggingFace repo ID in model_path, which ships as mlabonne/NeuralLlama-3-8B-Instruct-abliterated. First run downloads the weights, so budget for an 8B model's worth of disk and a slow first load.
• Local Safetensors - point at a .safetensors file via llm_safetensors_file instead.

Then the usual LLM dials: temperature (0.6 default), top_p (0.75), top_k (40), max_new_tokens (512), repetition_penalty (1.0). For video prompts you mostly want to leave these alone - the system prompt is doing the real work. Lower temperature for more consistent rewrites, raise it if every output comes back identical.

The one dependency trap

This node is the reason the pack README has a whole "Important Note About hpsv3" section. FunPack's LLM nodes need transformers >= 5.0. The optional hpsv3 package (used by the StoryMem keyframe quality filter) pins transformers==4.45.2. Install both and the Prompt Enhancer breaks. The two features are mutually exclusive on a single environment - pick one, or install hpsv3 with --no-build-isolation and understand what you're giving up.

Install

Part of ComfyUI-FunPack:

cd ComfyUI/custom_nodes
git clone https://github.com/olivv-cs/ComfyUI-FunPack
pip install -r requirements.txt

or ComfyUI Manager → search "ComfyUI-FunPack". Then confirm your transformers is ≥ 5.0 before first run.

Where it fits

Use it before a video refiner or directly in front of your sampler: enhancer output → conditioning → sample. The "Standalone" in the name is the clue - this is the LLM enhancer without needing the whole Studio rig. For the pipeline version, feed its output into FunPack Video Refiner V2, which treats the enhanced text as the "suggested prompt." It won't make a bad idea good, but it reliably turns a thin prompt into the kind of detailed description LTX and Wan actually respond to.