# Skill: Apply LoRA add-ons in ComfyUI for style and character consistency

> Lets an AI agent wire up Load LoRA nodes, set correct strength values, supply trigger words, and stack multiple LoRAs safely -- whenever a user wants a specific art style, consistent character, or targeted visual effect without retraining a model.

Source: sodigi·learn -- comfyui/lora-stili · https://sodigi.io/learn/comfyui/lora-stili

## When to use
- User wants a specific recognizable style (cyberpunk, ink, pixel art, etc.) that the base model does not produce reliably.
- User needs the same character or face across multiple generations.
- User asks where to get or how to install LoRA files.
- User reports that LoRA "is not working" or the style is too weak or too strong.

## Core rules
- A LoRA is a small patch file (megabytes, not gigabytes) that modifies one specific aspect of a model's output; it does not replace the model.
- LoRA must match the base model family it was trained on; a LoRA trained on one base will not work correctly with a different base.
- Place the LoRA file in `ComfyUI/models/loras/`; the checkpoint file goes in `models/checkpoints/` -- never mix them.
- The Load LoRA node is an adapter inserted between the model and the sampler; it passes through MODEL and CLIP outputs.
- Recommended starting strength: strength_model = 0.75, strength_clip = 0.75. Above 1.0 usually over-exposes the effect.
- Many LoRAs require trigger words in the prompt to activate; always read the author's description on Civitai for the required words.
- When stacking multiple LoRAs, chain Load LoRA nodes (output of one feeds the input of the next) and reduce each strength to 0.5-0.6.

## Procedure
1. Go to Civitai, filter by LoRA, and set the base model filter to match your checkpoint.
2. Open a LoRA page, read the example images, and note the trigger words from the description.
3. Download the safetensors file.
4. Place it in `ComfyUI/models/loras/`.
5. In ComfyUI, add a Load LoRA node. Connect its MODEL input from Load Checkpoint's MODEL output, and CLIP input from Load Checkpoint's CLIP output.
6. Connect Load LoRA's MODEL and CLIP outputs to KSampler's MODEL input and the CLIP Text Encode nodes respectively.
7. Set strength_model and strength_clip to 0.75 as a starting point.
8. Add the trigger words to the positive prompt and generate.
9. To stack a second LoRA, chain another Load LoRA node after the first; set each strength to 0.5-0.6.

## Ready-to-use prompt
```
I want to apply a LoRA style in ComfyUI. Help me:

1. Wire the Load LoRA node correctly between the checkpoint and the sampler.
2. Set a safe starting strength (0.75 on both sliders).
3. Find and add the trigger words for this LoRA to my positive prompt.
4. Understand why the LoRA seems to have no effect (likely: wrong base model, missing trigger word, or strength set to 0).

LoRA I want to use: [name or Civitai URL]
My base checkpoint family: [e.g., SDXL, SD 1.5, Flux]
My positive prompt so far: [your prompt]
```

## Pitfalls
- Forgetting to include the trigger word in the prompt -- the most common reason LoRA appears inactive.
- Setting strength above 1.0 hoping for a stronger effect -- typically produces an over-exposed, artifact-filled image.
- Using a LoRA trained on a different model family than the current checkpoint -- incompatible; effects will be wrong or absent.
- Placing the LoRA file in `models/checkpoints/` instead of `models/loras/`.
- Not refreshing or restarting ComfyUI after dropping the file in the folder.
- Stacking several LoRAs all at 0.8+ strength -- they conflict and degrade the image; lower each to 0.5-0.6 when combining.