ComfyUI LoRA Guide: How to Use, Stack, and Troubleshoot

What is LoRA?

LoRA (Low-Rank Adaptation) is a technique for fine-tuning AI models without retraining them from scratch. Instead of modifying all model parameters, LoRA adjusts a small subset — resulting in compact files (typically 10–200 MB) that can be loaded on top of any compatible base model.

Think of a base model like a versatile artist, and LoRAs as specialized training sessions — one might teach the artist a specific character design, another a particular art style, and another a lighting technique.

Why Use LoRA?

Getting Started

1. Download a LoRA

Find LoRAs on model sharing sites. Make sure the LoRA is compatible with your base model (e.g. SD1.5 LoRAs work with SD1.5 checkpoints, SDXL LoRAs with SDXL checkpoints).

2. Install the File

Place the downloaded .safetensors file in:

ComfyUI/models/loras/

3. Add the Node

In your workflow, add a Load LoRA node between the Load Checkpoint and the rest of your pipeline. Connect:

• model input → from Load Checkpoint's MODEL output
• clip input → from Load Checkpoint's CLIP output

The Load LoRA node outputs modified versions of both, which you then connect to your CLIP Text Encode and KSampler nodes.

Load LoRA Parameters

Tuning Strength

• 0.5–0.8 is a good starting range for most LoRAs
• Too high (>1.0) can cause artifacts or oversaturation
• Too low (<0.3) may have no visible effect
• Each LoRA has its own sweet spot — experiment to find it

Stacking Multiple LoRAs

You can chain multiple Load LoRA nodes in sequence. The output of one feeds into the input of the next:

Load Checkpoint → Load LoRA (style) → Load LoRA (character) → CLIP Text Encode → KSampler

When stacking:

• Lower individual strengths — start at 0.4–0.6 each to avoid conflicts
• Order can matter — try different arrangements if results aren't ideal
• Watch for conflicts — two LoRAs trained on similar concepts may interfere

Common LoRA Types

Tips for Best Results

• Use trigger words — many LoRAs require specific keywords in your prompt. Check the LoRA's description page for these
• Match base models — SD1.5 LoRAs don't work with SDXL and vice versa
• Test with simple prompts first — isolate the LoRA's effect before combining with complex prompts
• Refresh ComfyUI after adding new LoRA files if they don't appear in the node dropdown

Common Issues and Fixes

LoRA doesn't appear in the dropdown

• Make sure the file is in ComfyUI/models/loras/ (not a subfolder, unless ComfyUI is configured for recursive scanning)
• Refresh ComfyUI (F5) or restart it after adding new files
• Check the file extension — it should be .safetensors (not .ckpt or .pt)

LoRA has no visible effect

• Strength too low — increase strength_model to 0.7–1.0
• Missing trigger words — many LoRAs require specific keywords in your prompt. Check the model page on Civitai for trigger words
• Wrong base model — SD1.5 LoRAs don't work on SDXL checkpoints and vice versa

Output looks "deep fried" or oversaturated

• Strength too high — reduce strength_model to 0.5–0.7
• LoRA conflict — if stacking multiple LoRAs, lower each one's strength
• cfg too high — reduce the KSampler cfg value to 6–8

Error: "LoRA key not found" or similar

• Version mismatch — the LoRA was trained for a different model architecture
• Corrupted download — re-download the LoRA file

Next Steps

• Text to Image — Master the base workflow before adding LoRAs
• Image to Image — Combine LoRAs with reference images
• Upscale Guide — Enhance LoRA-generated images

Source References

• ComfyUI LoRA example
• LoraLoader node documentation
• ComfyUI models concept