IAMCCS Cine PromptRelay Timeline

ComfyUI-PromptRelay is the community's answer to one of LTX's core frustrations: the model takes a single global prompt, but long videos want "this happens, then that happens, at these moments." PromptRelay splits your video latent into segments and applies a different local prompt to each one, so the model gets a timeline instead of a one-liner. The catch is the format - you have to produce a list of segment lengths and a matching list of local prompts, and hand-building those so they add up to the exact 8n+1 frame count is exactly the kind of arithmetic that ends with a 400-frame latent and a 397-frame prompt list.

IAMCCS Cine PromptRelay Timeline is that math, in node form. You describe the timeline in human terms; it returns local_prompts, segment_lengths, and max_frames that slot straight into a PromptRelay encode node.

How it works

You give it:

• global_prompt - the scene-level prompt (passed through unchanged).
• timeline_data - either JSON (segments/rows/timeline with seconds + prompt + optional camera) or pipe lines: seconds | local prompt | camera note.
• duration_seconds and frame_rate - the shot.
• ltx_round_mode - defaults to up_8n_plus_1, i.e. round the total frame count up to the LTX-legal 8n+1 form.
• fallback_segments - if the timeline is empty, it fabricates that many evenly-spaced generic "cinematic beat n" segments so the graph still validates.

Then it converts seconds to proportional frame lengths that sum exactly to max_frames (any rounding remainder is dumped onto the last segment so you never get a dangling frame), joins the local prompts with |, and hands you the pieces. Camera notes get appended to their segment's prompt automatically.

The outputs that matter

• local_prompts - pipe-separated, one per segment.
• segment_lengths - comma-separated frame counts summing to max_frames.
• max_frames - the rounded total; wire it to your latent builder.
• global_prompt - passthrough.

The report helpfully includes a connect_to field pointing at PromptRelayEncodeTimeline or PromptRelayEncode.

Install

It's in IAMCCS-nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/IAMCCS/IAMCCS-nodes.git

or Manager → search IAMCCS. README floor: ComfyUI ≥ 0.3.0, Python ≥ 3.12, Torch ≥ 2.8. To actually encode, you need ComfyUI-PromptRelay in custom_nodes - this node only produces the data.

Where people get burned

• It produces data, not conditioning. Forgetting the second pack is the classic failure: this node runs, prints a lovely report, and your video ignores it because there's no PromptRelay encode node in the graph.
• Frame-sum drift. The node guarantees lengths add up to max_frames, but only if you feed that same max_frames into your latent builder. If your latent is built from a different rounding (say nearest vs up), the lengths won't match the latent again - which is exactly the mismatch PromptRelay complains about.
• Camera notes that repeat. A camera note already in the prompt text gets appended anyway (the code checks case-insensitively and skips it - but only for exact substring matches). Keep notes short and distinct.
• Zero search impressions, one-author tooling. The Cine family is documented mostly through tooltips and report outputs. For learning PromptRelay itself, the standalone ComfyUI-PromptRelay docs are better; this node is for the IAMCCS workflow specifically.