⏱ Dialogue Timer - stop the mumbling

A tiny ComfyUI node that sets your clip length from the words your characters actually speak.

Talking-video models hate silence. Make a 10 second clip with 4 seconds of dialogue and the model fills the gap with mumbling and gibberish - then finally says your lines. Make it too short and the lines get rushed or cut off.

Dialogue Timer reads your prompt, counts only the spoken words (not the scene description), and sets the clip length to fit. You write the scene, it works out the seconds.

What it counts

• MiniMax H3's speech tag - <d>[English] Follow the wind, live free.</d>

• Anything in "quotes" - she says: "Here is your coffee."

• No quotes at all? It takes the words after says / asks / replies / whispers / shouts... up to the end of the sentence.

Camera notes, <Picture 1> tags and descriptions are ignored.

The math

• spoken words ÷ 2.5 words per second, + 1.5 seconds of breathing room

• kept between 5 and 15 seconds, rounded to the nearest half second

• 10 spoken words = 5.5 seconds

• wrote too much? It tells you: "60 spoken words needs 25.5s - TOO MUCH for one clip, split it"

Every number is a setting you can change, and there is an AUTO / MANUAL switch if you want to set the length yourself.

Install

• Download the zip and unzip it

• Put the ComfyUI-DialogueTimer folder into ComfyUI/custom_nodes/

• Restart ComfyUI

• No extra Python packages needed

How to use it

• Double-click the canvas, type Dialogue Timer

• Connect your prompt text to its prompt input

• Connect its seconds output to wherever your workflow sets the clip length

• Optional: connect report to a Preview Any node to read what it decided

The zip includes an instant demo workflow - no models needed, just drag it in and press Run. A full guide is in the README.

Workflow wants frames instead of seconds? Multiply seconds by your frame rate with any math node.

Three tips for clean dialogue

• Put speech in "quotes" - it helps the model as much as the timer

• Keep character names inside the quotes only - a name repeated in the description tends to get spoken out loud

• End a line with an action - "...then he is quiet and takes a sip" - so the model knows the talking is over

Built for MiniMax H3, but it is model-independent - it works with any workflow that takes a clip length. MIT licence, do what you like with it.