LoRA: Style and Character Fine-Tuning

What a LoRA Adds to Your Model

A LoRA (Low-Rank Adaptation) is a small set of weights trained on top of a base checkpoint. It does not replace the checkpoint — it modifies it at runtime by pushing the model's outputs toward a specific style, character, object, or concept. LoRAs are typically 50 MB to 500 MB compared to a 5–10 GB checkpoint, which makes them fast to swap and easy to mix.

Common LoRA types:

• Style LoRAs: teach the model to render in a specific artistic style — watercolor, cel shading, photorealism, pencil sketch, a specific illustrator's aesthetic
• Character LoRAs: make the model generate a specific fictional character, celebrity, or created persona with consistent appearance
• Concept LoRAs: add a specific object, clothing item, or environment the base model cannot generate well
• Face LoRAs: improve portrait accuracy for a specific person

LoRA files go in your ComfyUI/models/loras/ folder. Download them from CivitAI (filter by model type = LoRA) or Hugging Face. Always verify the base model the LoRA was trained on — an SDXL LoRA will not work correctly on an SD 1.5 checkpoint.

Adding the Load LoRA Node

To add a LoRA to any workflow:

1. Double-click the canvas and search for "Load LoRA" to add the node
2. Disconnect the MODEL and CLIP wires coming out of your Load Checkpoint node
3. Connect Load Checkpoint → MODEL to Load LoRA → model input
4. Connect Load Checkpoint → CLIP to Load LoRA → clip input
5. Connect Load LoRA → MODEL to your KSampler → model input
6. Connect Load LoRA → CLIP to your CLIP Text Encode nodes → clip input
7. In the Load LoRA node, select your LoRA file from the lora_name dropdown

The Load LoRA node sits in the middle of the MODEL and CLIP flow. It reads the checkpoint outputs, applies the LoRA weights, and passes the modified model and CLIP forward to the rest of the workflow.

Setting LoRA Weights

The Load LoRA node has two weight sliders: strength_model and strength_clip (both default 1.0, range 0.0 to 3.0+).

• strength_model controls how strongly the LoRA changes image generation
• strength_clip controls how strongly the LoRA changes text interpretation

In most cases, set both to the same value. Recommended starting weights by LoRA type:

• Style LoRA: 0.4–0.7 — keep it lower than you think; a style LoRA at 1.0 often produces artifacts or looks cartoonish
• Character LoRA: 0.7–1.0 — needs higher weight to maintain consistent appearance
• Face / portrait LoRA: 0.6–0.9 — go higher for stronger identity, lower to preserve variation
• Concept / object LoRA: 0.5–0.8 — depends heavily on the specific LoRA

Generate a test image at weight 0.6, then 0.8, then 1.0. Compare the three. The sweet spot is usually where the style appears clearly without artifacts dominating the output.

Stacking Two or Three LoRAs

You can apply multiple LoRAs to the same generation by chaining Load LoRA nodes. The first LoRA's MODEL and CLIP outputs connect to the second LoRA's inputs:

• Load Checkpoint → MODEL/CLIP → Load LoRA 1 (style) → Load LoRA 2 (character) → KSampler and CLIP Text Encode

When stacking, reduce each LoRA's weight compared to what you'd use solo. A common starting point for two LoRAs is 0.5 + 0.5. The combined effect should add up to no more than about 1.0–1.2 total influence. Going over this budget produces artifacts and incoherent outputs. Experiment in 0.1 increments.

Three LoRAs is the practical maximum for most workflows. Beyond that, the competing fine-tune signals usually degrade output quality.

Trigger Words and CivitAI Notes

Many LoRAs require a specific trigger word in your prompt to activate the fine-tuned concept. This word was used during training to associate the LoRA's effect with that token. Without the trigger word, a character LoRA may do nothing; with it, the character appears consistently.

Trigger words are listed on the LoRA's CivitAI page under the "Trigger Words" section. Examples: "ohwx person" for a person LoRA, "cyberpunk city" for an environment LoRA. Include the trigger word early in your positive prompt.

CivitAI's LoRA pages also show example images generated with recommended weights and prompts. These are the fastest way to understand how a LoRA behaves before you invest time building a workflow around it.