🥁 LoraGeek Continuum Beatbox - write the story, get the whole H3 Continuum timeline

Type your story as plain beats. Press one button. Your H3 Continuum story prompt is written for you, and the sampler's Chunks and Seconds are set to fit.

H3 Continuum makes long MiniMax H3 videos out of short chunks that flow into each other - same people, same voices, same room. It is amazing, but writing its timeline by hand is fiddly: every chunk needs its own [0-5s] section, the time marks must sit on their own lines, the chunk count has to match, and one slip makes the whole story play twice.

The Continuum Beatbox does all of that for you.

Why I built this

Continuum finally lets MiniMax H3 tell a story - a scene that runs for 30 seconds, a minute, more - without the faces, the room or the voices changing every time a new clip starts. That is the thing we have all been fighting.

But the first stories I tried went wrong in ways that had nothing to do with the model. The time marks sat on the same line as the text, so Continuum quietly ignored the whole timeline and played the entire story in every chunk - he said his line twice, in two different voices. The chunk count didn't match the story. A spoken line got split across a chunk boundary. A description of the dress in the prompt overrode the reference picture.

Every one of those is a formatting mistake, and a tool should never let you make them. So now the Beatbox writes the timeline, and you just write the story.

Write it like you'd tell it

5s the woman walks into the cafe and waves at the man
The man stands up, smiles and says: You made it!
She hugs him and says: "Sorry I'm late, the traffic was crazy."
sound: a coffee machine hisses
5s they sit down and the waiter sets two cups of coffee on the table
He raises his cup and says: To finally catching up.
4s they clink their cups and both laugh

Press GENERATE PROMPT → 6 chunks x 5 s = 30 seconds of story, written into Continuum's story box, sampler set, ready to run.

What it does for you

• Packs your beats into Continuum's chunks - and never cuts a spoken line in half at a chunk boundary

• Times every beat - speech from its words, actions from a default, or your own: start a line with 5s

• Writes a header that Continuum repeats on every chunk - the cast, the place, the look and the soundscape stay the same all the way through

• Sets the Continuum Sampler - Chunks, Seconds per chunk and Timeline mode, automatically

• Handles who's talking - every spoken line is tied to that person's reference (The man (@R3) says: ...) so the words come out of the right mouth. She says goes to the last woman you mentioned, and every line ends with "then her lips close" so the talking stops cleanly

• Use @R tags in your beats - @R3 says: I love you., @R3: I love you. or @R1 walks by: I'll be right with you. all work - always use says, a colon or quotes for speech

• Natural acting - the header keeps everyone looking at each other with small natural movements between beats, instead of freezing and staring into the lens

• Sound per beat - sound: a door closes puts that sound exactly in that beat's chunk

• Quiet chunks don't babble - a filler line (She: Mm-hm.) keeps H3 from inventing mumbling

• Shows you the plan - 6 beats -> 6 chunks x 5 s = 30 s, and which beat landed in which chunk

🎥 The Beatbox Camera - a field of buttons

Click your look, shot size, angle, lens, lighting and camera move (push in, pull out, pan, arc, tracking, POV...). Every choice goes into the header of every chunk. Click the i next to any section for a plain-English explanation. Buttons with a star come straight from the official H3 prompt guide.

You can SEE the prompt

Nothing is hidden. The finished timeline lands in Continuum's own story box - read it, change any word, and run.

The full prompt it writes for that story

This is exactly what lands in Continuum's story box - nothing trimmed:

@R1 is the woman. @R2 is the man.
Everyone keeps the same face, hair, clothes and body as in their reference pictures. They are absorbed in the moment and look at each other.
Live-action, cinematic, a medium shot, lit by soft natural window light.
A sunny corner cafe in the afternoon, big windows, plants on the tables.
The camera pushes in at slow speed.
overall_soundscape: quiet cafe chatter, cups clinking on saucers, a coffee machine hissing now and then.
The actions happen in the order described, with small natural movements in between: glances, smiles, shifting in their seats.

[0-5s]
The woman walks into the cafe and waves at the man. The woman (@R1) says: <d>[English] Mm-hm.</d> Then her lips close.

[5-10s]
The man stands up, smiles. The man (@R2) says: <d>[English] You made it!</d> Then his lips close.

[10-15s]
She hugs him. The woman (@R1) says: <d>[English] Sorry I'm late, the traffic was crazy.</d> Then her lips close. Sound: a coffee machine hisses.

[15-20s]
They sit down and the waiter sets two cups of coffee on the table. The woman (@R1) says: <d>[English] Mm-hm.</d> Then her lips close.

[20-25s]
He raises his cup. The man (@R2) says: <d>[English] To finally catching up.</d> Then his lips close.

[25-30s]
They clink their cups and both laugh. The woman (@R1) says: <d>[English] Mm-hm.</d> Then her lips close.

Everything above the first [0-5s] is the header. Continuum repeats it on every chunk, which is why the cafe, the window light, the slow push-in and the coffee-shop sounds hold steady from the first second to the last.

Install

• Download the zip and unzip it

• Put the ComfyUI-ComplexPromptGenerator folder into ComfyUI/custom_nodes/ (already have my Beatbox? Replace the old folder - this pack includes it)

• Restart ComfyUI, then hard-refresh the browser page (Ctrl+Shift+R)

• No extra Python packages. Costs no VRAM - it's a template, not an AI

• You need H3 Continuum installed

How to use it

• Open your H3 Continuum workflow

• Double-click the canvas, search Continuum Beatbox - the camera panel appears beside it. Nothing to wire

• Cast: @R1 = the woman, @R2 = the man - names only, your reference pictures decide how they look

• Write your beats, click camera buttons, press GENERATE PROMPT, run

• A ready demo workflow (Continuum's official V3.9 workflow + the Beatbox) is in the zip

FAQ

Do I need the original Beatbox? No - this pack includes it. If you already have it, replace the old folder with this one.

Does it use an AI model or VRAM? No. It is a template written in plain JavaScript. It runs instantly and adds nothing to your render.

Does it work without H3 Continuum? The Continuum Beatbox is made for Continuum's sampler. For single H3 clips use the original Beatbox (Complex Prompt Generator) in the same pack.

Does it work with reference pictures and a start picture? Yes. Put who is in each Reference slot in Cast (@R1 = the woman). Your start picture goes in Continuum's First Frame as usual.

My characters stare into the camera - why? Give every beat a small natural action (leans in, touches his hand, laughs) and pick Two-shot in the camera panel. The Beatbox already tells H3 they look at each other.

Why names only in Cast? Anything you write there is repeated on every chunk and wins over the picture. Write "green dress" and she will wear green even if her reference picture shows red.

Can I edit the prompt after it is written? Yes - it is plain text in Continuum's story box. Change a word and run. Press GENERATE again only when you want it rebuilt from the beats.

How long can a story be? Continuum does up to 16 chunks per run - 80 seconds at 5 s per chunk, about 2.5 minutes at 10 s. For longer films, write one scene per run and join them.

Which models? Whatever your Continuum workflow uses. A start picture plus references works best on a hybrid FL2VA/Ref2VA H3 model.

Tips

• Keep chunks at 5-10 seconds

• Use real room sounds in the soundscape - H3 renders music instruments poorly

• Continuum does up to 16 chunks per run - split longer stories into scenes

• Turn on Continuum's Progress / Run Storage to redo one bad chunk without re-rendering the rest

Also in the pack: the original Beatbox (Complex Prompt Generator) for single H3 clips.

MIT licence. Made by LoraGeek. Continuum is by its own author - go give it a like.