Boyo Prompt Relay Encode

A video prompt is usually one sentence for the whole clip, and the model does its best to make that work. Prompt Relay fixes the actual limitation: it lets different parts of the video be driven by different prompts, with smooth temporal transitions between them. This node is a faithful port of kijai's ComfyUI-PromptRelay into Boyonodes - same logic, attributed, renamed so both packs can coexist - with the classic combo: a global_prompt that anchors the whole video plus local_prompts that take over in sequence.

So if you want a video where a character walks through a forest, then a city, then a desert - without any cuts - you write three local prompts and let the gating blend between them.

How it works

You feed it a model, a clip, and an empty latent video whose dimensions define the temporal budget. The core idea: each local prompt is encoded normally, then masked with a Gaussian curve across time. Prompt A is at full strength for the first segment, fades out as prompt B fades in, and so on. The node then patches the model to apply that temporally-varying conditioning during sampling - that's why it outputs a model as well as the positive conditioning. You take both outputs downstream to your sampler.

The inputs that matter:

• global_prompt - conditions the entire video. This is where you anchor persistent stuff: character identity, scene context, camera style.
• local_prompts - ordered, one per segment, separated by |. "walk through a forest|walk through a city|walk through a desert".
• segment_lengths - comma-separated pixel-space frame counts per segment. Leave empty and it auto-distributes evenly. If you give lengths, they're converted to latent lengths internally (largest-remainder method), so you don't have to do latent math yourself.
• epsilon - the Gaussian penalty decay. Default 0.001 gives sharp segment boundaries (that's the paper default). Below ~0.1 you get crisp transitions; raise it to 0.5+ for soft, gradual cross-fades. This is the single knob most people actually touch.

Where it fits

Anything where the content of a video changes over its runtime: scene-transition animations, music-video-style cuts, storyboard clips where each scene is its own prompt. The node is model-agnostic at the conditioning level, so it works in Wan and LTX workflows alike - the same temporal-relay idea the LTX Director tooling popularised.

Use a standard CLIPTextEncode for the negative prompt; the relay only handles positives.

Installing it

Ships with Boyonodes:

cd ComfyUI/custom_nodes
git clone https://github.com/DragonDiffusionbyBoyo/Boyonodes

Restart ComfyUI, or install "Boyonodes" via ComfyUI Manager. Requires the pack's prompt_relay.py and patches.py modules, which ship with the pack - no external deps beyond what ComfyUI already has.

Gotchas

• local_prompts must contain at least one prompt or the node raises. The | separator is mandatory; don't use commas.
• If your video quality collapses at segment edges, your epsilon is too low - soften it before touching anything else.
• Because it patches the model, the patched model output is the one you must use in the sampler; plugging the un-patched model in silently ignores the relay.

The most interesting node in the pack for video people. If you've been fighting "one prompt per clip" constraints, this is the lever.