FunPack CLIP Vision Output Combine

ComfyUI is full of nodes that accept one CLIP_VISION_OUTPUT and exactly one. So the moment a workflow has two different CLIP Vision encodes - say, one for a reference image and one for a keyframe - you're stuck with a wire that can't connect. FunPack CLIP Vision Output Combine is the small adapter that fixes that: it takes up to four CLIP_VISION_OUTPUT inputs and merges them into a single output of the same type.

How it works

The method input (default mean) decides how the tensor fields get combined. Your options are mean, median, maximum, and minimum. The rule is simple: every tensor field being merged must have the same shape, and non-tensor fields are copied through from the first input. So this is a pointwise merge, not a clever learned fusion - which is exactly what you want from an adapter node.

Why would you combine CLIP Vision outputs at all? In FunPack's own world, the Refiner and Studio take a clip_vision_output as advisory image context - "here's what the source looks like, don't drift from it." If you're passing multiple reference frames, merging their vision embeddings first keeps the downstream node reading one clean description instead of you having to pick a favorite. It's also handy when two CLIP Vision models disagree about a scene: averaging them can smooth out the noise.

The inputs that matter

Only one is required:

• clip_vision_output1 (CLIP_VISION_OUTPUT) - always needed.
• clip_vision_output2/3/4 (CLIP_VISION_OUTPUT, optional) - add up to three more.
• method (enum) - mean is the safe default; median is the robust choice if one of your sources is noisy; maximum/minimum are for when you want to bias toward the strongest or weakest response.

Output is a single clip_vision_output. Simple, one type in, one type out, no surprises.

Install

It comes with the whole pack:

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

or ComfyUI Manager → search "ComfyUI-FunPack". It's a pure tensor merge - no dependencies beyond what the pack already needs, no model downloads, no GPU cost beyond a few tensors being averaged.

The honest take

This is a boring node and that's fine. It won't improve your generation quality by itself, it won't teach the model anything - it just removes a "I can't connect this wire" moment. That's a real service, because every multi-reference video workflow eventually hits it. Use mean unless you have a reason not to, keep your inputs the same shape, and treat it as the plumbing it is.