FunPack Advisor LLM

FunPack Video Refiner V2 has an "advisor": a text model that looks at your rating, your prompt, and the run history, and suggests what to fix - "motion keywords are underrepresented," that sort of thing. FunPack Advisor LLM is the node that supplies that brain. It loads any HuggingFace CausalLM and exposes it as an advisor_clip output, which you wire straight into the Refiner V2 (or Studio) advisor_clip input. The name is the whole story: this is a model loader for one specific job.

How it works

One model_path string and one dtype choice. The path is either a HuggingFace repo ID or an absolute local path to a model directory. The default is huihui-ai/Huihui-Qwen3-8B-abliterated-v2 - an 8B Qwen instruct model, "abliterated" meaning the refusal behavior was trained out. That's the author's choice for an uncensored setup (the pack has no content restrictions), but you can drop in any CausalLM you like.

dtype (default bfloat16) is the precision: bfloat16 recommended for CUDA, float32 for CPU. If you're on a low-VRAM card, float16 or float32 are your levers - though honestly, an 8B model is the part of this pack where VRAM gets real. First load downloads the weights and caches the model, so the first run is slow and later runs reuse it. That cache is shared with Studio's built-in advisor, so the model loads only once even if both are in the workflow.

Output is a single advisor_clip (CLIP type). That's it - wire it into advisor_clip and the Refiner's advisor feature lights up (set advisor_mode to anything but Off on the Refiner).

Why you'd use the standalone at all

If you live in Studio, the Advisor tab does this internally. The standalone node exists for the graph-based workflow - where you want the same advisor model loaded explicitly, or you want to swap models without touching Studio's popup. It's also the cleaner setup if you're already using Studio with an external advisor_clip input: the input always overrides Studio's internal setting.

The dependency trap

This is an LLM node, so the pack-wide rule applies hard: it needs transformers >= 5.0, and the optional hpsv3 package pins transformers to 4.45.2 and will break it. Since hpsv3 is only used by the StoryMem keyframe quality filter, you usually shouldn't install it at all.

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". No additional packages beyond the pack requirements - but be ready for the model download on first run.

The honest take

This node is a bridge: without an advisor-capable model connected, the Refiner V2's advisor is inert, and the Refiner is the best feature in this pack. Don't skip it because the default is an 8B model. If VRAM is tight, load a smaller instruct model locally and point model_path at it - the advisor's job is analysis and repair, not generation, and a smaller model doing that job well beats a giant one OOMing.