Apply Trained LoRA

What it is

ApplyTrainedLora is the small, deliberately dumb node that closes the loop on this pack's ten training nodes: it takes a lora_path string - the output of any RealtimeLoraTrainer, SDXLLoraTrainer, or Musubi trainer in the pack - and patches it onto your model. It exists because a freshly trained LoRA lives on disk with a generated filename, not as an entry in ComfyUI's standard LoRA dropdown, so the usual LoraLoader node can't see it yet without a restart. This node reads the path directly instead, which is what makes "train it and immediately test it in the same workflow" actually work.

Everything about it is intentionally plain: no per-block controls, no analysis, no scheduling. If you want any of that, this pack's Selective LoRA Loaders and Analyzer+Selective Loader V2 nodes are the ones to reach for. This node's whole job is being the fastest path from "I just trained something" to "does it look right."

How it works

It's a standard LoRA-application node with one twist: instead of picking a filename from a dropdown populated at ComfyUI startup, it accepts a live file path as an input, which lets it pick up a LoRA that was written to disk moments earlier in the same run. Connect a trainer's lora_path output straight into this node's lora_path input and the two form a tight train-then-test loop without touching the filesystem yourself.

Inputs and outputs that matter

• model - the base model to patch, same as any LoRA loader.
• lora_path - a string, not a dropdown selection. This is the field that takes a trainer node's output directly.
• strength - a float from -10 to 10 (default 1), the usual LoRA weight dial. Note the range is wider here than the pack's Selective Loaders (-5 to 5), presumably because a freshly trained LoRA sometimes needs pushing harder than a published one to show its effect clearly during testing.

Outputs are model (the patched model, wire it into your sampler as usual), plus lora_name and lora_path as pass-through strings - handy if you want to log or display which file actually got applied, since the filename alone won't tell you much when you're iterating through several training runs.

Installing it

It ships with the rest of the pack. ComfyUI Manager (search "Realtime LoRA Trainer") or:

cd ComfyUI/custom_nodes
git clone https://github.com/ShootTheSound/comfyUI-Realtime-Lora

Restart ComfyUI. This node needs no external training backend and works the moment the pack is installed - it's a loader, not a trainer, so none of the sd-scripts/Musubi/AI-Toolkit setup applies here.

Common issues

Nothing happens when you connect a trainer's output. Double-check the trainer actually completed - if keep_lora was left off on the trainer node, or the run errored partway, the path this node receives may point at a file that was never written. Check the trainer node's console output first.

Strength 1.0 looks too weak or too strong. This is normal variance between training runs rather than a bug in this node - a LoRA trained at low rank on a small dataset can need a different strength than one from a bigger run, even on the same architecture. Nudge the strength slider rather than assuming your training settings were wrong.

Wrong model connected. Like every LoRA loader, this one silently produces garbage if the base model input doesn't match the architecture the LoRA was trained on - a Z-Image LoRA on an SDXL model, for instance, won't error, it'll just look wrong. Match your model input to whichever trainer node produced the LoRA.