How to Use a LoRA in ComfyUI

Add a LoRA to your workflow to push a model toward a specific style or subject, and learn how to balance its strength against the base model.

7 minIntermediate

A LoRA is a small file that nudges a base model toward a particular style, character, or concept without retraining the whole thing. In ComfyUI you apply one by inserting a Load LoRA node between your checkpoint and the rest of the graph. This guide shows the wiring and how to tune its strength.

What you need

A working text-to-image workflow (SD 1.5, SDXL, or Flux)
A LoRA file that matches your base model's family
Any trigger words the LoRA's page lists

Step 1: Put the LoRA file in the right folder

Drop the .safetensors LoRA into models/loras. A LoRA trained on SDXL only works with SDXL checkpoints, and a Flux LoRA only works with Flux. Mixing families gives garbled or no results.

Step 2: Add a Load LoRA node

Right-click the canvas, choose Add Node > loaders > Load LoRA. This node takes a model and a clip input and outputs a modified model and clip.

Step 3: Wire it between checkpoint and sampler

Connect the checkpoint's MODEL and CLIP outputs into the Load LoRA node, then route the LoRA node's MODEL and CLIP outputs onward to the sampler and the text encoders. The LoRA must sit in the middle of that chain to take effect.

ComfyUI - LoRA in the chain

[ Load Checkpoint ] --MODEL--> [ Load LoRA ] --MODEL--> [ KSampler ]

\--CLIP----> \--CLIP----> [ CLIP Text Encode ]

Load LoRA

lora_name: my_style.safetensors

strength_model: 0.80

strength_clip: 0.80

Checkpoint feeds the LoRA, the LoRA feeds everything downstream.

Step 4: Add the trigger words to your prompt

Many LoRAs are trained with specific trigger words that activate the concept. Check the LoRA's download page and add those words to your positive prompt, otherwise the LoRA may barely show up.

Tune the strength

Start strength_model around 0.7 to 0.8. Too high and the LoRA overpowers the prompt and distorts anatomy; too low and you barely see its effect. Adjust in 0.1 steps until the style is present but the image still holds together.

Step 5: Stack multiple LoRAs if needed

To combine, say, a style LoRA and a character LoRA, chain a second Load LoRA node after the first. Each node modifies the model and clip from the one before it. Keep the combined strengths modest so they do not fight each other.

example positive prompt

a portrait of a knight in a misty forest, dramatic lighting,
mystyle, intricate armor, cinematic

Result: the same base model now renders in your chosen style. Save the workflow once you find a strength that works so you can reuse it without rewiring.