Lora Loader (Searge)

This is Searge's own LoRA loader, and functionally it does exactly what you expect: it applies a LoRA to your model and CLIP at whatever strength you set. If you've used ComfyUI's built-in Load LoRA node, you already know this one - same inputs, same outputs, same idea. It applies a LoRA's weight patches to the diffusion model and the text encoder so you can add a style, a character, or a concept on top of your base checkpoint.

The thing to flag right away: it's in the pack's _deprecated_ category. It exists so that older Searge workflows keep loading, not because it does anything the standard ComfyUI loader doesn't. In a fresh graph, reach for the built-in.

How it works

A LoRA is a small set of low-rank weight adjustments trained on top of a base model. Loading it means merging those adjustments into the model at runtime, scaled by a strength value. This node does that for both halves that a LoRA can affect - the UNet (strength_model) and the text encoder (strength_clip) - and hands you back the patched pair.

The inputs and outputs that matter

• model (MODEL) and clip (CLIP) - the checkpoint's model and CLIP going in. You chain these through if you're stacking multiple LoRAs.
• lora_name - the LoRA file to load, from your models/loras folder.
• strength_model (default 1.0) - how strongly it affects the image. This is the one you'll actually tune; 0.6-0.9 is a common range for not overpowering the base.
• strength_clip (default 1.0) - how strongly it affects prompt interpretation. Often left equal to strength_model, sometimes dropped lower.

Outputs are the patched MODEL and CLIP. To stack LoRAs, feed these into the next loader's inputs.

How to install it

Manager: search SeargeSDXL, install, restart. Manual, opencv first:

python -m pip install opencv-python
cd ComfyUI/custom_nodes
git clone https://github.com/SeargeDP/SeargeSDXL.git

Restart ComfyUI. LoRA files go in ComfyUI/models/loras - the README, for example, links the SDXL offset-noise example LoRA if you want to try adding contrast.

Common issues & troubleshooting

LoRA does nothing / trained for the wrong base. SDXL LoRAs only work on SDXL checkpoints, and there are family caveats within SDXL - a LoRA trained on a Pony-line model may misbehave on an Illustrious-line one and vice versa. If a LoRA looks dead, check it actually matches your checkpoint's family before touching strengths.

Overcooked images. Strength 1.0 is often too much, especially when stacking several LoRAs. Pull strength_model down to 0.6-0.8. The effects compound, so three LoRAs at full strength will fight each other.

Should you use this at all? For a plain LoRA load, the built-in ComfyUI Load LoRA node is the equivalent and isn't marked deprecated - prefer it in new work. Searge's v4.x workflow also does multi-LoRA (up to five) through its own data-stream system, which is the intended path inside that workflow. This standalone loader is legacy glue.