Prompt Relay Advanced Options
Fine-tune how your Prompt Relay segments soften at the seams
- relay_options
If Prompt Relay Encode is the steering wheel, this is the fine adjustment you make after you've driven the car. Prompt Relay Advanced Options (C2C) produces a RELAY_OPTIONS bundle you feed into the encoder's optional relay_options socket to tune how per-segment prompts transition across time. It's entirely optional - the encoders run fine without it - but it's where you go when the default segment blending isn't cutting it.
What it tunes
The concept it tunes: Prompt Relay works by giving each temporal segment of a video its own prompt and patching the model's cross-attention so each segment's influence rises and falls across frames. Two things shape how that feels - how far out the falloff starts, and how sharply it cuts. This node exposes those dials separately for the video stream and the audio stream.
The three knobs
Three knobs, each with a video and audio twin:
video_strength/audio_strength(default 1) - multiplier on the temporal penalty. 0 disables segmentation entirely, which makes the whole video one flat prompt. The author's guidance is worth quoting: most useful in 0–1 to soften boundaries, and >1 "saturates quickly at the default epsilon - raise epsilon to ~0.1 to make >1 meaningful." In practice: push below 1 for gradual crossfades, and if you want genuinely hard cuts, that's mostly epsilon's job.video_window_scale/audio_window_scale(default 1) - scales the flat anchor zone in the middle of each segment (default roughlyL/2 - 2frames where nothing fades). Below 1 starts the falloff sooner, so segments overlap more; above 1 widens the rigid, fully-owned middle and leaves less transition room.audio_epsilon(default 0) - the LTX audio stream's own epsilon. 0 means "inherit from the encoder," so you only touch it if your audio conditioning behaves differently from the video stream.
The output, and when to bother
One output: relay_options, a single opaque bundle type. Wire it into PromptRelayEncodeC2C (any backend) - the two deprecated encoders also accept it. It's the kind of node that makes sense only after you've seen a default relay and thought "the cut between scene two and three is too soft" - and then it's exactly the knob you needed.
The real-world tuning advice: for most single-scene-video work you can ignore this node entirely. Reach for it when you have (a) long clips where segments blend too early, so raise video_window_scale; (b) transitions that feel abrupt, so drop video_strength into 0.5–0.8 and/or soften with a higher epsilon on the encoder; or (c) audio conditioning that fights the video cuts, so set audio_epsilon explicitly instead of inheriting.
Install
There's no model download, no dependency beyond the pack itself - install is clone Code2Collapse/ComfyUI-CustomNodePacks into custom_nodes or ComfyUI Manager → "CustomNodePacks". It's a tuning surface, not a feature, and knowing that is half the battle.
Inputs (5)
| Name | Type | Default | Description |
|---|---|---|---|
| video_strength | FLOAT | 1.000–10 | Multiplier on the temporal penalty. 0 disables segmentation. Most useful in 0–1 to soften boundaries; >1 saturates quickly at the default epsilon — raise epsilon to ~0.1 to make >1 meaningful. |
| video_window_scale | FLOAT | 1.000–4 | Scales the flat anchor zone (default L/2 - 2 frames). <1 starts falloff sooner; >1 widens the rigid zone. |
| audio_epsilon | FLOAT | 0.00000–0.99 | LTX audio stream epsilon. 0 = inherit from the encoder. |
| audio_strength | FLOAT | 1.000–10 | Multiplier on the temporal penalty. 0 disables segmentation. Most useful in 0–1 to soften boundaries; >1 saturates quickly at the default epsilon — raise epsilon to ~0.1 to make >1 meaningful. |
| audio_window_scale | FLOAT | 1.000–4 | Scales the flat anchor zone (default L/2 - 2 frames). <1 starts falloff sooner; >1 widens the rigid zone. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| relay_options | RELAY_OPTIONS | — |