🎬 Scene Builder - plain beats in, a working reference prompt out
A tiny ComfyUI node that writes your reference-to-video prompt in the structure the model actually needs.
Reference-to-video models (MiniMax H3 Ref2VA and friends) are picky about how a prompt is written. Get the structure slightly wrong and you get:
characters mumbling gibberish between their lines
a voice-over with nobody's lips moving
someone chanting a name five times
<Picture 3>pointing at the wrong picture because you switched a slot off
Scene Builder writes that structure for you. Say what each picture is once, then for every scene just list what happens, one thing per line:
she walks over carrying a coffee
she says: Here is your coffee. Careful, it is hot.
he says: Thank you. I really needed this today.
he takes a sip
...and it writes the whole prompt.
No requirements. None.
No models. No LoRAs. No extra Python packages. This is not a workflow you have to go hunting files for - it is a small text tool that drops into your workflow. The demo uses only built-in ComfyUI nodes and opens without a single red box.
What it does for you
Senses which picture and voice slots are switched on. Bypass a picture and it drops out of the prompt, and every
<Picture N>renumbers itself the way the model really counts themGroups pictures of the same subject. Label them the natural way - the woman, the woman side shot, the woman on a bicycle - and it writes "...are the same woman seen from different angles". Any number of pictures, any description. Props and places too
Writes dialogue so lips actually move - a named speaker, "says once", tied to the right
<Audio N>voiceKeeps spoken lines back to back and moves actions out of the gap, so nobody invents mumbling (switch it off to keep your exact order)
Gives the model real sounds to make - these models always generate audio; tell them what the room sounds like and they stop inventing speech
Change one box to move the whole scene - the bathroom becomes the kitchen, everything else stays
Never rewrites your words - your beats go in exactly as you typed them
ON / OFF switch - OFF passes your own hand-written prompt through untouched, so adding the node never locks your workflow
Install
Download the zip and unzip it
Put the ComfyUI-SceneBuilder folder into ComfyUI/custom_nodes/
Restart ComfyUI
How to wire it
Double-click the canvas, type Scene Builder
Connect each Load Image to a picture socket as well as to your reference node. Same for Load Audio and the voice sockets. That is how it senses what is switched on
Connect its prompt output to your reference node's prompt
Optional: connect report and prompt to Preview Any nodes to read what it did
It tells you when something is off: "Picture slot 4 is ON but its box is empty", "the man speaks but has no voice slot".
Works even better with the Dialogue Timer
My other node sets the clip length from the spoken words. Connect the Scene Builder's action_seconds to the timer's breathing_room and the clip length follows the scene you wrote. There is also a combined download with both nodes in one folder.
Full guide in the README. MIT licence, do what you like with it.
Description
First release. List what happens in plain words, one beat per line, and it writes the full reference-to-video prompt: senses which picture and voice slots are on and renumbers them, groups pictures of the same subject, writes dialogue so lips move, keeps spoken lines back to back so nobody mumbles, and adds real room sounds. ON / OFF switch. No models, LoRAs or extra packages needed. Includes an instant demo using only built-in nodes.