Merge LoRA Model

Merge LoRA Model is the "bake it in" option in PG Nodes' LoRA toolkit. It takes a MODEL, pulls a LoRA out of a LORA_BRIDGE bundle, and merges the LoRA weights directly into the UNet/model path - destructively. The output is a patched MODEL, not a runtime toggle. Strength is locked at merge time; there's no fading, no switching, no un-apply.

That's the trade: it's fast and dead simple, and you give up the flexibility the hook path gives you. In the same pack, the hook route (Create LoRA Model Hook → Set Model Hooks) patches the model live so you can keyframe strength over the denoising run. This node is for when you know you want a LoRA on the model path at a fixed strength and don't need the theater.

How it works

Inputs are model (the MODEL from your checkpoint or UNet loader), lora_bridge (the bundle from the LoRA Bridge Loader), which_lora (a dropdown picking slot lora_1 through lora_4), and strength_model (a float from −100 to 100, default 1.0). It grabs the chosen slot from the bridge and calls ComfyUI's own load_lora_for_models with strength_model applied to the model path and 0.0 to the clip path - hence "model only." The text encoder is untouched, which is the difference between this node and the both-paths merge.

A strength of 0 returns the model unchanged. Negative strengths are allowed, and some people genuinely use them to push a LoRA in reverse - though honestly, most of the time that's a sign you picked the wrong LoRA.

Two behaviors worth knowing:

• It silently passes the original model through if the bridge slot is empty or the merge fails (with a [PgMergeLoraModelOnly] apply error in the console). No crash, which is friendly, but also easy to miss - check the log if nothing looks different.
• This is the "old-school" baked-in pattern from the pack's own connection-pattern docs: LoRA Bridge Loader → Merge LoRA Model → downstream. One-line wiring, zero runtime overhead.

The inputs that matter

• model - your MODEL. Wire straight from a loader.
• lora_bridge - from LoRA Bridge Loader.
• which_lora - which of the bridge's four slots to merge.
• strength_model - how hard to apply it. 1.0 is standard; the range goes to ±100 for the bold.

Output: a single patched MODEL that feeds your sampler (via a sampler/guider setup). There's no CLIP output - that's what the Clip-only and both merge nodes are for.

Install

Part of the PG Nodes pack:

ComfyUI Manager → search "PG Nodes" → Install → restart

Manual: clone https://github.com/GizmoR13/PG-Nodes into ComfyUI/custom_nodes (folder named PG-nodes), restart, and it's under PG/Lora. No extra dependencies beyond ComfyUI itself.

Common issues

• Output looks like the input. Check the console for the apply error line, and confirm the bridge slot actually has a LoRA (the Bridge Loader logs load errors). Silent passthrough on empty slots is by design.
• "Why can't I toggle it off?" Because you baked it. That's this node's job. If you need per-step control, use the hook path - keyframes are the feature for that.
• LoRA not in the bridge dropdown. New files appear after a ComfyUI restart; the list is built at import.

If you're coming from A1111 or an older ComfyUI habit, this node feels like home - it's exactly the merge behavior you already knew, just routed through a bridge. Choose it when you want predictable and simple; choose hooks when you want control.