Boyo Prompt Relay LoRA Gate

Temporal prompting is great, but prompts aren't the only thing that changes mid-video. Sometimes the style needs to change - a character LoRA for one scene, a different aesthetic for the next, a speed LoRA for one segment only. BoyoPromptRelayLoraGate is the extension to Boyo Prompt Relay that makes a LoRA apply to exactly one temporal segment of a video, fading in and out with the same Gaussian gating that Prompt Relay uses for prompts. It's a genuinely distinctive feature - someone in a community thread even pointed people at this exact node when asked about per-segment LoRA gating.

How it works

It takes a model, a lora_name (dropdown of your installed LoRAs), a segment_index, and a strength. The mechanism: instead of applying the LoRA at full strength across the whole generation, the node weights the LoRA's effect with a Gaussian curve centred on your chosen segment. At the segment's centre you get the full strength (default 0.8); as you move away in time, the influence fades toward zero. The output is the patched model - one gate per segment, chained.

The critical input is segment_index: it's zero-based and must match the segment order in your Boyo Prompt Relay Encode node. Segment 0 is the first local prompt, segment 1 the second, and so on. The node doesn't verify the match - it trusts you - so if your prompt order and your gate order drift apart, you'll get the right LoRA on the wrong scene and it'll be subtle enough to confuse you.

Where it fits

The intended pattern is chaining: one gate per segment, each targeting its own segment_index, each fed the previous gate's model output. Bypass the gates you don't need (the node's docs explicitly recommend bypassing unused ones rather than deleting them). So a three-scene video with a different LoRA per scene is three gates in series:

Model → Gate(segment 0, LoRA A) → Gate(segment 1, LoRA B) → Gate(segment 2, LoRA C) → Sampler

Combined with Prompt Relay Encode's per-segment prompts, this is how you get a video that's prompt-switched and style-switched per scene in a single continuous generation - no cuts, no stitching.

Installing it

Ships with Boyonodes (DragonDiffusionbyBoyo):

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

Restart ComfyUI or install "Boyonodes" via ComfyUI Manager. No extra dependencies; the gating logic lives in the pack's prompt_relay.py / patches.py, which come with the pack.

Gotchas

• segment_index is zero-based and manually matched. There's no wire connecting the gate to the relay's segment list - a mismatch just silently gates the wrong time window.
• strength maxes at 2.0 but the default 0.8 is a sane starting point; pushing toward 2.0 can overpower the segment and fight the base model.
• It only patches what you feed it - if you use the un-gated model anywhere in parallel, that path ignores the LoRA entirely.

One node, one segment, one LoRA - chain them and a whole video's style arc is just a series of gates. There isn't much else out there doing this cleanly in ComfyUI.