Load Caption

Load Caption is the least impressive node in the ComfyUI-ThinkSound pack, and that's fine - it's supposed to be. In the ThinkSound scheme of things, it's the "what is this scene" input: a short, plain-English title that gets paired with a much longer Chain-of-Thought description from its sibling node, Load CoT Description. Together those two text inputs condition Alibaba's ThinkSound audio model, and this one supplies the coarse caption while the other one supplies the detailed sound design.

You'll reach for it every time you build the pack's workflow, because the main ThinkSound node requires a title input and there's no built-in alternative. It's one of those nodes you wire once and mostly stop thinking about.

How it works

Mechanically it's a glorified text passthrough. The node takes your string and returns it unchanged as a PROMPT-typed output named title. From there the ThinkSound node writes it into the first column of a cot.csv - the short caption that sits alongside the longer CoT description the model conditions on. No models, no tokenizers, no API. It literally cannot fail on its own.

The one input that matters

• text - a multiline string, default "Printer Printing". Whatever you type here is what the model sees as the scene's name.

That's it. One input, one output. The output title (type PROMPT) wires straight into the title port on the ThinkSound node.

A small note on style: keep this short. ThinkSound conditions on both the title and the detailed description, and the description is where the actual sound design lives. "Rain on a tin roof" works. A wall of text here is just redundant with what Load CoT Description should be doing.

Installing it

Load Caption ships in the ComfyUI-ThinkSound pack, so you get all four nodes at once. Search "ComfyUI-ThinkSound" in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Yuan-ManX/ComfyUI-ThinkSound.git

The pack needs the ThinkSound weights in ckpts/ (from Hugging Face liuhuadai/ThinkSound or ModelScope) and its requirements are a fully pinned environment - install those in a dedicated venv, not your main ComfyUI one, or you risk torch and numpy getting replaced. Full details live in the ThinkSound node article.

Gotchas

The loaders themselves are bulletproof, so the only thing to know is upstream of them: the main ThinkSound node is currently a non-functional stub as shipped - it calls its generation pipeline without consuming it, and the scripts it needs aren't in the repo. Until that's fixed, don't be surprised that a perfectly-wired graph produces no audio. The text nodes are doing their job; the engine is the part that's half-built.