📦 Dialogue Tools - Scene Builder + Dialogue Timer in one download
Two small ComfyUI nodes for talking-video workflows. One install. Use either, or both together.
No requirements. None.
No models. No LoRAs. No extra Python packages. These are not workflows you have to go hunting files for. They are small text tools that drop into your workflow, whatever model you use. The demo uses only built-in ComfyUI nodes and opens without a single red box.
🎬 Scene Builder
Say what each picture is once, then for every scene just list what happens, one thing per line. It writes the full reference-to-video prompt in the structure the model needs:
senses which picture and voice slots are switched on and renumbers every
<Picture N>correctlygroups pictures of the same subject - the woman, the woman side shot, the woman on a bicycle
writes dialogue so lips actually move - named speaker, "says once", tied to the right voice
keeps spoken lines back to back so nobody mumbles in the gap
gives the model real room sounds instead of silence it fills with gibberish
never rewrites your words, and has an ON / OFF switch so it never locks your workflow
⏱ Dialogue Timer
Counts only the spoken words in the prompt and sets the clip length to fit. Too long and the model mumbles to fill the silence; too short and lines get cut off. Reads MiniMax H3 <d> speech tags, "quotes", or plain "she says..." sentences.
Together
Connect the Scene Builder's prompt to the timer's prompt, and its action_seconds to the timer's breathing_room. Write the scene - the prompt and the clip length sort themselves out.
Install
Download the zip and unzip it
Put the ComfyUI-DialogueTools folder into ComfyUI/custom_nodes/
Restart ComfyUI
Already have my standalone Dialogue Timer? No problem - it is the same node. You can delete the standalone folder.
Both nodes are also available as separate downloads on my profile if you only want one. Full guides are in the zip. MIT licence, do what you like with it.
Description
Fix: actions written between two spoken lines are no longer moved to the end of the clip. They now stay where you wrote them and happen during the dialogue. Before, characters had nothing to do while talking, and the model filled the time with its own ideas (like undressing when nobody asked). New 'hold the scene' switch, on by default: tells the model nobody does anything your beats don't describe and clothing stays as it is. Turn it off if you want the model to improvise. Thanks to the users who reported this.
FAQ
Comments (8)
Hey this looks promising! Is there a way to see what the nodes are actually generating when it sends it to the video model?
For now all I see is:
5 seconds (7 spoken words needs 4.3s)
1 pictures ON, 0 voices ON | 2 spoken lines (7 words), 2 actions
Pictures: <Picture 1> = a woman
Room for actions handed to the timer: 1.5 s
! Actions that sat between two spoken lines were moved after the last line, so nobody mumbles in the gap.
Oh thats the Report tool. On the other box that is the one that builds the prompt. Run that one. It will give you the entire prompt. LMK
Happy the play with this a bit but as the other person alluded to, the "! Actions that sat between two spoken lines were moved after the last line, so nobody mumbles in the gap." is a major issue as the characters seem free to do whatever for 15 seconds during the dialog. With no prompting or anything even leaning in that direction one of my female characters keeps removing the top of her dress to get her breasts out. Turns out I did not know she was an exhibitionist at heart ;)
Ill look into this when I get home. Thanks for the heads up!
Thanks for reporting this - you were right, and it's fixed in v1.0.1. The old 'close the gaps' option pushed every action written between two lines to the very end of the clip, so while people were talking the prompt gave them nothing to do - and the model filled the gap with its own ideas (hence the surprise exhibitionist!). In v1.0.1 those actions stay where you wrote them and happen during the dialogue, and a new 'hold the scene' switch (on by default) tells the model nobody does anything your beats don't describe and clothing stays as it is. Grab v1.0.1 and let me know how it behaves for you.
Which finetune are you using? Seems like an interesting side-effect 😂
@FrogOnStilts Ha - honestly I think my tool gets the blame on this one, not the finetune! The old version left the characters standing around with nothing to do while they talked, so the model got creative. For what it's worth, I test on stock MiniMax H3 Ref2VA with the official 8-step ref turbo. The nodes themselves don't need any model - they only write the prompt, so they work with whatever you run.
@LORAGEEK I'm just happy to be to a point where I can be useful. Spent a year constantly breaking things that everyone else seemed smart enough to run without issue. Now I'm 50/50 whether it is my fault or not.