LTXV Prompt Relay Encode (Scheduled)

The name is doing a lot of work, so let's unpack it: LTXV Prompt Relay Encode (Scheduled) is the glue node of Jinx138's Temporal Director, the node that makes your prompt changes and your LoRA envelopes share one clock. It takes the segment prompts from LTXV Schedule Sync, encodes them the way Kijai's ComfyUI-PromptRelay would, and - the important part - emits a relay_timing signal so every scheduled Time-Gated LoRA can match its transition softness to the prompt transitions. No timeline UI, no extra bookkeeping: the schedule lives in segment_data, and this node just rides it.

If you've done LTX 2.3 for a while you know the pain this pack exists to kill: you want a claymation look for the middle third of a clip and realism everywhere else, and your only "solution" was generating separate clips and stitching them. PromptRelay solved the prompt side of that. This node extends it so the LoRA side moves with the prompt side, instead of you hand-timing two unrelated schedules.

How it works

Under the hood it adapts Kijai's PromptRelay rather than reimplementing it: the pack finds prompt_relay.py and patches.py inside your ComfyUI-PromptRelay folder, builds the token ranges for your global + local prompts, converts the visible-frame segment lengths from Schedule Sync into latent-frame lengths, and patches the model with PromptRelay's mask function.

The piece that makes the Director work is relay_timing. The epsilon you set here (default 0.001) becomes a sigma shoulder via sigma = 1 / ln(1/epsilon), converted to an edge-softness measured in frames. Smaller epsilon means a sharper prompt transition; larger means a softer blend. Every Time-Gated LoRA that connects relay_timing derives its default curve shoulder from this same value, so prompts and LoRAs ease in and out at the same moment.

Inputs and outputs that matter

You mostly set three things:

• global_prompt - the steady base prompt for the whole video (the "realistic" in the claymation example).
• segment_data - straight from LTXV Schedule Sync. This carries the local (per-segment) prompts and the exact segment lengths.
• epsilon - transition softness; start at the 0.001 default and only raise it when prompt cuts feel abrupt.

The latent input must be the same video-only LTX latent the rest of the Director uses - its shape is what the scheduling math reads. There's also an optional relay_options socket for Kijai's Prompt Relay Advanced Options if you use those.

Outputs: model (the patched model, chained into your TimeGate nodes then the sampler), positive (the scheduled conditioning for your sampler), schedule_info (a debug string), and relay_timing - which you wire to every scheduled Time-Gated LoRA in the graph.

Installing and the one thing that bites

This is the one node in the pack that hard-requires ComfyUI-PromptRelay. If you see a runtime error about the PromptRelay module not being adapted, that's what it means. Install it, enable it, restart ComfyUI. The classic Time-Gated LoRA nodes don't need it; this one does.

Otherwise it's the standard pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/Jinx138/ComfyUI-LTXV-TimeGated-LoRA.git

Restart ComfyUI, hard-refresh the browser, and grab it under "LTXV/Temporal Director". No pip packages to install, no model downloads - you supply the LTX 2.3 checkpoint and your visual LoRAs. And one author-flagged tip from the release thread: trigger-dependent style LoRAs work best when the trigger word lives inside the active segment's prompt, not the global one.