Load LoRA (Bypass) (For debugging)

Let's get the honest part out of the way: Load LoRA (Bypass) is not a node you reach for to make better images. Its display name even says so - "(For debugging)". If you just want to slap a LoRA onto a model, use the regular Load LoRA node and don't look back. This one exists for a narrow technical situation, and it's genuinely useful there.

The situation is training. A normal LoRA loader takes your LoRA and patches the weights - it bakes the adapter's effect into the model's parameters, permanently, for the lifetime of that node instance. That's fast and fine for inference. But when you're training a LoRA, or running a model whose weights are offloaded to CPU and you don't want to drag a patched copy around, permanently modifying weights is exactly what you don't want. That's what the bypass mode fixes.

How it works

Instead of modifying base weights, bypass mode injects the LoRA computation into the forward pass. The source docstring puts it in one clean formula:

output = base_forward(x) + lora_path(x)

The base model runs untouched, and the LoRA's contribution is computed and added on top. Non-adapter LoRA bits (bias diffs, weight diffs) still get applied as regular patches, but the adapter part rides along in the forward pass rather than being folded into the weights. If you set both strengths to zero, the node short-circuits and hands you back the original model and CLIP unchanged.

Inputs and outputs

Same shape as the regular Load LoRA, minus the mystery:

• model - the diffusion model the LoRA applies to
• clip - the CLIP model the LoRA applies to
• lora_name - the LoRA file from models/loras/
• strength_model - how hard the LoRA hits the diffusion model; negative values work
• strength_clip - same knob for CLIP

Outputs are the modified MODEL and CLIP, wired into the sampler and text encoder just like any other LoRA chain.

Where it lives

It's a recent addition - the debug LoRA nodes were consolidated into their own file in January 2026 and marked experimental. It ships with core ComfyUI, so there's nothing to install; if you don't see it, update.

The trap

The trap is treating it as a drop-in replacement for the normal loader and wondering why things behave differently. Bypass mode keeps your base weights pristine, which is the point - but that also means you're paying a forward-pass overhead on every inference, and it exists for training and offload scenarios, not for your everyday generation pipeline. There's a model-only sibling (Load LoRA (Bypass, Model Only)) if you don't care about CLIP at all. For everyone else: enjoy knowing it's there, and keep using Load LoRA.