PM Apply LoRA

The merge is done, the merged LoRABundle is sitting there - now what? If you want to use it without saving it as a file, this is the node: it patches your checkpoint's model and CLIP with the merged LoRA, exactly like the standard LoraLoader does, and hands you a working MODEL and CLIP to run through the sampler.

It's the "test the merge before you commit it to disk" step. The alternative - PM LoRA Save - writes a .safetensors file you'd then load with a normal loader. Apply is faster feedback: wire it in, sample, and decide if the merge is worth saving at all. You'll spend most of your merge experiments here, and only save the winners.

How it works

The merged LoRA bundle carries both UNet-side and CLIP-side weights (assuming you left merge_clip on in the merger). This node applies the UNet patches to model and the CLIP patches to clip, mirroring ComfyUI's standard two-output LoraLoader behavior. If the merger excluded CLIP layers, only the model gets patched - the CLIP output just passes through.

The inputs that matter

• model - your diffusion model (from the checkpoint loader or anything downstream).
• clip - the matching CLIP model.
• lora - the LoRABundle output from PM LoRA Merger (or anywhere else a bundle appears, like PM LoRA Select).

Two outputs, MODEL and CLIP, which feed the sampler as usual.

Installing

Part of the LoRA Power-Merger pack. ComfyUI Manager (search "LoRA Power-Merger") or:

cd ComfyUI/custom_nodes
git clone https://github.com/larsupb/LoRA-Merger-ComfyUI
cd LoRA-Merger-ComfyUI
pip install -r requirements.txt

Restart ComfyUI; the node shows as PM Apply LoRA. Dependencies: PyTorch, lxml, mergekit.

Common issues

The most common confusion is treating this like a standard LoRA loader and expecting a strength widget - there isn't one. The strength was baked in at merge time (via lambda_ and the stack strengths), so to change the effect you re-merge, not re-apply. If your merged LoRA has no visible effect, check the merger's lambda_ and merge_clip settings rather than this node. And match the model: applying an SDXL merged LoRA to an SD1.5 checkpoint will fail on key mismatches - the merged bundle isn't a magic cross-architecture adapter.