H3 Relay · Apply Model LoRA

H3RelayModelLoRA - displayed as H3 Relay · Apply Model LoRA - is where you'd put a style or character LoRA onto an H3 Relay model bundle. You feed it a H3_RELAY_MODEL wire, pick a LoRA, set a strength, and get back a patched H3_RELAY_MODEL wire. It's the bundle-aware version of ComfyUI's LoraLoaderModelOnly, and mechanically that's exactly what it expands into.

The interesting part is what happens around the LoRA. H3 Relay's model bundles carry a hidden cache fingerprint - the cache_tag that records exactly which checkpoint, LoRAs, and strengths went into the model chain. When this node patches a bundle, it recomputes that fingerprint from the parent's tag plus the new LoRA name and strength. That's the mechanism behind the pack's smart cache invalidation: swap a LoRA or change a strength and only the artifacts derived from the patched chain go stale. Everything generated before your edit stays valid. The model, meanwhile, is loaded once and fanned out - patchers stay lazy, so a cache hit doesn't re-materialize checkpoint weights on the GPU.

The inputs that matter

• model (H3_RELAY_MODEL) - any bundle: H3 or LTX. The node preserves the bundle's kind, so the pack always knows whether you're patching the raw-generation model or the LTX finishing model.
• lora_name - dropdown from your models/loras folder.
• strength (default 1, range −100 to 100) - the LoRA weight. The range looks absurd until you realize negative strengths are a legitimate way to push away from a LoRA's direction. For normal use, 0.5–1.0 is the working zone.

Output: model (H3_RELAY_MODEL), wired onward to the next LoRA, an Attention Backend node, or Generate Shot / LTX 2× Enhance.

How to chain it

You can stack several of these on the same bundle, one after another, before the shot node. Because each patch updates the fingerprint, the cache correctly treats "model + LoRA A" and "model + LoRA A + LoRA B" as different model chains - changing the stack rerolls only the derived work, never the raw accepted shots that don't depend on it.

One caution that isn't this node's fault: LoRAs are architecture-specific. A LoRA trained for vanilla LTX will misbehave on an H3-model bundle, and vice versa. The kind tag keeps H3 Relay honest about which family a bundle belongs to; it won't stop a mismatched LoRA from producing garbage, so check your LoRA's target before you load it.

Install

Shared pack install: ComfyUI Manager → search H3 Relay → install → restart, or git clone https://github.com/akatz-ai/h3-relay into custom_nodes. No extra deps beyond the models the pack already needs; the LoRA files themselves go in models/loras and are, like everything weight-shaped, covered by their own license rather than the pack's GPL-3.0 source license.