LoRA — Custom Styles

LoRA (Low-Rank Adaptation) is a small file that teaches an AI model a specific visual style. Think of it like a color grade LUT — the LUT doesn’t change what your camera captures, it transforms the look consistently. You bring the content, the LoRA brings the style.

When to Use LoRA

• Brand consistency across a campaign — same illustration style across 20 social clips without re-prompting every time
• Character continuity — keep the same character look across multiple scenes or episodes
• Matching a client’s mood board — lock a specific photographic or illustration aesthetic from reference images
• Recreating a film stock look — match the grain, color, and lighting of a specific era or camera

Prompt vs. LoRA

What prompts do	What LoRA does
Describes content (subject, scene, action)	Defines visual style (colors, texture, aesthetic)
Changes every generation	Stays consistent across generations
Anyone can write	Trained on specific reference images
Flexible, different each time	Locked visual language

They work together: the prompt controls what appears, the LoRA controls how it looks. If you just need “cinematic golden hour” — prompt it. LoRAs are for styles you can’t describe in words.

Finding Style URLs

Style files are hosted on two main platforms:

• CivitAI — Browse the LoRA section. Copy the direct download link for the .safetensors file.
• Hugging Face — Search for LoRA models. Copy the direct URL to the .safetensors file in the model’s “Files” tab.

The URL must point directly to the .safetensors file — not to a model page or preview. Make sure the LoRA is compatible with your base model — a Flux 1 LoRA won’t work on Flux 2 Pro.

Using a Custom Style

1. Select a model with the custom styles badge in model search
2. In the generate tab, paste the style URL into the LoRA URL field
3. Set the LoRA Scale to control how strongly the style is applied

The hint bar above the URL field will confirm when your style is active.

LoRA Scale

The scale controls how much influence the style has on the output:

Value	Effect
0.5	Subtle influence — the style is present but doesn’t overpower
0.7	Recommended starting point — noticeable style without overriding the base model
0.9	Strong style application — clear visual transformation
1.0	Full strength — may introduce artifacts on some models
>1.0	Amplified — likely to cause visual artifacts or distortion. Avoid.

Start at 0.7 and adjust by 0.1 increments. Every style behaves differently, so experiment to find the right balance.

Stacking Multiple LoRAs

Some models accept multiple LoRA URLs. Two LoRAs at 0.5 each can work, but keep the combined weight under 1.5. Three or more LoRAs almost always produces visual chaos — use one at a time, two maximum.

Trigger Words

Some styles require specific words in your prompt to activate. These are called trigger words.

Check the style’s page on CivitAI or Hugging Face for instructions. If trigger words are listed, include them in your prompt — otherwise the style may not take effect.

Example: A style might require the word ohwx in the prompt to activate its effect.

Common Issues

Style has no visible effect

• Make sure the URL points to the actual .safetensors file, not a page
• Check if the style requires trigger words in the prompt
• Try increasing the scale to 0.8 or 0.9

Output looks distorted or has artifacts

• Lower the scale — try 0.5 or 0.6
• Some styles are trained for specific resolutions or aspect ratios — match the style’s recommended settings

URL rejected by the model

• The URL must be publicly accessible (not behind a login wall)
• Use the direct download link, not a page URL
• CivitAI URLs may require you to accept the model’s terms first

Quick Answers

Is a LoRA the same as a fine-tuned model? Not exactly. A fine-tuned model is completely retrained. A LoRA is a lightweight add-on that modifies an existing model’s behavior — smaller, faster to swap, and can be combined.

Can I train my own LoRA? Yes, but it requires a separate training process outside modelBridge. You provide 10–30 reference images and use a training service. modelBridge supports using the result — not creating it.

Do LoRAs work with video models? Some video models support LoRAs, but it’s more common with image models (especially Flux). Check whether the model’s advanced settings include a LoRA field.