Load CoT Description

This is the node that makes ThinkSound different from the other video-to-audio models. Load Caption gives the scene a name; Load CoT Description is where you hand the model the plan - a step-by-step, chain-of-thought write-up of the sounds that should happen and roughly when. Alibaba's ThinkSound conditions on that detailed reasoning, not just a caption, and this node is the text input that feeds it.

Think of it as writing a cue sheet for a sound designer instead of tagging a clip with a one-liner. That's a genuinely different workflow from MMAudio and the other caption-in/audio-out models, and it's why the default text in this node is worth studying before you overwrite it.

How it works

Like its sibling, it's a pure text passthrough: your string goes in, and comes out unchanged as a PROMPT output named description. The main ThinkSound node drops it into a cot.csv under a caption_cot column, which the conditioning stack - T5 text embeddings, MetaCLIP, SyncNet - turns into the guidance for the diffusion run. No computation happens here; the node just gets your text to where it matters.

The input, and how to write it well

• text - a multiline string. The default is a genuine mini-masterclass in CoT prompting:

Generate a continuous printer printing sound with periodic beeps and paper movement, plus a cat pawing at the machine. Add subtle ambient room noise for authenticity, keeping the focus on printing, beeps, and the cat's interaction.

Notice the structure: it names the primary sound first, adds secondary events ("periodic beeps", "paper movement"), introduces a second subject, throws in ambience, and finishes by stating what not to lose focus on. That's the template. When you write your own, aim for that - concrete sounds, a rough timeline, and a priority. "Sad ambience" will get you mush; "low rumbling engine with a door slam halfway through" gets you something usable.

One output, description (type PROMPT), which plugs into the description port on the ThinkSound node alongside the title from Load Caption.

Installing it

It ships in ComfyUI-ThinkSound with the rest of the pack. Search "ComfyUI-ThinkSound" in ComfyUI Manager, or:

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

Then the shared pack setup: clone the weights into ckpts/ (Hugging Face liuhuadai/ThinkSound or ModelScope), and install requirements in a separate environment - they're pinned to exact torch/numpy versions that will stomp your ComfyUI install. See the ThinkSound node article for the full rundown.

Gotchas

Two things worth knowing. First, keep the title and the description on the same page - a caption that contradicts the CoT plan just muddies the conditioning. Second, and more important: right now this node feeds a pipeline that doesn't run. The main ThinkSound node is a stub as shipped, so a correctly-wired description won't produce audio until the pack's generation wiring is fixed. Your prompting skills aren't the problem; the engine is. Worth keeping an eye on, because when it's working this input is the difference between generic whooshes and audio that actually matches the footage.