Load LoRA Model

Most LoRA loaders hand you a file dropdown and a strength slider and call it a day. Load LoRA Model is the odd one out: it takes a LORA_MODEL wire - an already-loaded LoRA object, not a file name - and it has a bypass switch. If you're generating images, you probably don't want this node. If you're training LoRAs in ComfyUI, or you run distilled models on offloaded weights, this is where the interesting stuff happens. The node description is upfront about it: this is the training-path loader.

How it works

Four inputs:

• model - the diffusion model the LoRA applies to.
• lora - a LORA_MODEL object. That type is produced by the training nodes in the same core module (the ones that output trained LoRA weights alongside a loss map), not by a file picker. If you want to load a LoRA from disk, you want the ordinary Load LoRA node, which internally reads the file and patches. This node is the next stage: "here's a LoRA that already exists as an object, apply it to this model."
• strength_model - how strongly to apply it, default 1.0, negative allowed.
• bypass - a boolean that changes how the patch happens. Default off means the standard path: the LoRA's low-rank matrices get folded into the base weights. Flip it on and the node uses bypass mode instead.

That bypass flag is the reason this node exists, and it's worth understanding because it's the same trick the "for debugging" LoRA loaders use. Bypass mode computes output = base_forward(x) + lora_path(x) - the LoRA computation is injected during the forward pass rather than baked into the weights. The base model's weights are never modified. The tooltip spells out why you'd do this: it's useful for training, and for when model weights are offloaded (e.g., quantized models where you can't directly modify weights). Regular apply modifies weights; bypass doesn't.

When you'd actually use it

Realistically, this node appears in ComfyUI's new in-app training workflows - train a LoRA, then immediately test it on a model without leaving the graph. The bypass mode also quietly matters for distilled models: distilled checkpoints are often loaded as LoRAs stacked on a dev model, and there are workflows where bypass application is the correct or only safe choice. The debugging-focused sibling Load LoRA (Bypass, Model Only) is the same concept in a stripped-down form for when you don't have the training pipeline around.

Gotchas

• No file dropdown, no trigger-word magic. If you were expecting to browse models/loras, you're at the wrong node. This one consumes an object; it doesn't discover files.
• Bypass isn't free. Injecting the LoRA in the forward pass instead of folding it into weights can change speed and, with some samplers, behavior. The community's Wan 2.2 experience is a cautionary tale: bypassing a distilled LoRA outright needs more steps and CFG > 1, because the distilled behavior was part of the weight.
• Experimental badge. Like the whole training node family, this is core but marked experimental - expect it to move as the training story matures.

It ships with ComfyUI core. If you're not training, file it under "good to know exists." If you are, this is the node that lets you validate a freshly trained LoRA against the model in the same session.