Muse Collective LTX Timeline V2.5

V2.5 is V2 with one addition, and that addition is aimed at a very specific workflow: letting a local LLM write your segment prompts. If you've been hand-typing eight different prompts into an LTX timeline for a long-form video, you've probably wished for exactly this - a way to pipe text from an LLM node straight into the director's segments without copy-pasting.

Muse Collective is upfront that V2.5 is a fork of V2 with V2 left unmodified. Everything V2 gives you - the full timeline director, Seed Hunt, chunking, lipsync, layered audio - is still there. What V2.5 adds is segment_override_1..4, four optional text inputs that override the timeline's segment prompt text when connected and non-empty.

How it works

The mechanism is boring in the best way. Each segment_override_N input maps to a timeline segment (segment 0 through 3, despite the 1-based input names). Wire a string into it and, if it's non-empty, that text wins over whatever's typed in the segment's prompt box. Leave it unwired or empty and the timeline prompt is used as normal. There's no magic - it's a prioritization layer on top of the existing per-segment prompt system.

The natural setup is an LLM node (Ollama, LM Studio, QwenVL - anything that outputs text) whose job is to expand a scene description into LTX-flavored segment prompts, then wire its outputs into the override inputs. LTX likes long, detailed prompts, and a decent LLM is genuinely better at writing those than most humans typing quickly. The pack's own roadmap goes further down this path - V10 pairs with a dedicated MusePromptSplitter node that parses up to eight labeled prompts out of one LLM text block.

Inputs and outputs

The four overrides are the story: segment_override_1 through segment_override_4, all STRING with multiline. Everything else is V2's familiar panel - seed_hunt and the use_seed_hunt_N toggles, stage1_steps/stage2_steps/stage2_denoise, carry_frames, the audio toggles, and the rest.

Outputs match V2: last_chunk_frames, audio, stage1_frames, plus the four seed_hunt_preview_N / seed_hunt_audio_N pairs and reference_image.

Installing it

Identical to the rest of the pack:

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

Restart, pip install av torchaudio soundfile, and the full LTX 2.3 model stack. No extra dependencies for the overrides themselves.

Gotchas

• Four segments, not five. Despite the 1..4 naming, these map to timeline segments 0-3. For more segments you're into V10 territory (eight overrides).
• The override only takes effect when connected and non-empty - an empty override is silently ignored, which is actually the right behavior for a defaults system, but it can confuse you the first time you wire something in and see nothing change. Check the connection, not the string.
• This is one of the WIP modules the pack loads with a try/except. If the node doesn't appear in your node list at all, the pack silently skipped it - update the pack, and check the console log.

The honest verdict: if you're already running a local LLM and generating multi-segment LTX timelines, this is a small but real quality-of-life win. If you're not, it's nothing to rush for - V2 will treat you fine. But V2.5 is where the "let the LLM direct" pipeline starts to look genuinely good.