Muse Prompt Splitter

This is the glue node for the "let an LLM direct your timeline" pipeline. You feed it one block of LLM-generated text and it splits it into up to eight separate segment prompts and eight separate per-segment durations, using labeled markers in the text. It exists to solve the specific awkwardness of wiring an LLM's multi-part output into MuseDirectorSamplerV10's segment_override_1..8 inputs without copy-pasting or hand-splitting.

It's a pure text utility - no models, no GPU, no dependencies beyond Python's regex module. That makes it one of the friendliest nodes in this pack to get working.

How it works

Two independent regex passes over the input:

• PROMPT_N: markers become prompt_N outputs - PROMPT_1: starts the text for segment 1, and parsing stops at the next PROMPT_N: or DURATION_N: marker.
• DURATION_N: markers become duration_N outputs - parsed by a separate regex so duration text can never leak into the prompt outputs.

That separation is the whole design point. A DURATION_2: line sitting directly under PROMPT_2: is captured as a duration, not appended to the prompt text, because the prompt regex explicitly stops at DURATION_N: markers too. If you've ever watched a model's duration estimate get mangled into the prompt it was supposed to describe, you know why this matters.

The workflow it pairs with: an LLM node (Ollama, LM Studio, QwenVL - anything that returns text) with a system prompt instructing it to expand a scene description into labeled LTX 2.3 segment prompts, then:

PROMPT_1: A woman walks into a cafe, [SOUNDS] door chime, coffee machine hiss
DURATION_1: 12
PROMPT_2: She sits down, [SPEECH] I'll have the usual, please
DURATION_2: 8

Wire prompt_1..8 into V10's segment_override_1..8 inputs. The duration_N outputs are informational only - the node won't apply them to the timeline for you; read them and type them into the timeline by hand (or wire to a Show Text node).

Inputs and outputs

One input: llm_output (multiline STRING, forced as an input connection). Sixteen outputs: prompt_1..8 and duration_1..8, all STRING. Empty slots return empty strings, so downstream nodes always get the full set.

Installing it

Part of the same pack:

cd ComfyUI/custom_nodes
git clone https://github.com/muse-collective-26/muse-ltx-timeline

Restart. That's it - it shares the pack's av/torchaudio/soundfile requirements by virtue of living in the same folder, but the node itself uses only re.

Gotchas

• The LLM has to follow the format. No PROMPT_N: markers found → the node logs a warning and returns all empty strings. If your LLM drifts from the label format, fix the system prompt, not the node.
• Marker matching is case-insensitive (prompt_1: works), but the input needs the _N numbering - bare PROMPT: with no index is ignored.
• It's capped at 8; segments past that are silently dropped.

If you've already got an LLM writing your LTX prompts, this node is the missing seam between "model wrote me a whole script" and "the director needs per-segment text." Ten minutes to set up, and it removes the fiddliest manual step in the pipeline.