Load LoRA (Bypass, Model Only) (for debugging)

Let's be honest about this node up front: the display name ends with "(for debugging)", and that's the whole story. Load LoRA (Bypass, Model Only) is not how you add a character LoRA to your images. It's a specialized loader that applies a LoRA in bypass mode - meaning it doesn't modify the base model's weights at all. Instead it injects the LoRA computation during the forward pass: output = base_forward(x) + lora_path(x). You'd reach for it in training scenarios, when you're running offloaded or quantized weights that can't be modified in place, or when you're deliberately debugging whether a LoRA is the thing breaking your output. It's the model-only variant of the pair - the full version also patches a CLIP.

How it works

The inputs are identical to the normal model-only loader:

• model - the diffusion model the LoRA applies to.
• lora_name - the file from ComfyUI/models/loras.
• strength_model - default 1.0, negative allowed.

One output: the modified MODEL. The difference from the plain Load LoRA is purely how the patch is applied. Normal application folds the low-rank matrices into the checkpoint's weights once, at load. Bypass mode keeps a hook around that adds the LoRA's contribution every forward pass. Same file, same dial, different arithmetic - which is why it's perfect for the one thing that matters in debugging: it leaves the base model untouched, so a bypass-loaded LoRA can't corrupt weights that are shared or offloaded.

Why "for debugging"

The phrase in the name isn't decoration. ComfyUI ships this node (marked experimental) as part of its newer LoRA tooling, and the community pattern for it is narrow and technical. The most common real use is training: if you're fine-tuning a LoRA and want to preview its effect on a quantized model, bypass is the only safe way, because you literally cannot modify the weights of a GGUF-quantized model in place. The second use is diagnosis - stack a LoRA in bypass mode next to the same LoRA applied normally, compare, and you've isolated exactly what the weight-modification step does.

There's also a general lesson that transfers to everyone, not just trainers: bypassing a distilled LoRA changes behavior more than people expect. On Wan 2.2, for instance, bypassing the distilled LoRA outright produced blurry output until users raised steps and CFG above 1 - the distilled behavior lived in the weight, and removing it from the equation removed the model's shortcuts. If you're experimenting with bypass to save memory, budget for that.

Should you use it?

If your workflow is "load LoRA, generate, save," no - reach for Load LoRA (Model Only) and don't think twice. This node exists for the person who is actively training, debugging an offload problem, or A/B testing whether weight modification is the culprit. It's core, so there's nothing to install, but the "(for debugging)" label is the most honest display name in ComfyUI. Treat it as a scalpel, not a spoon.