Visual ToolsUpdated Aug 2026

Training custom visual models

Learn how to train custom LoRA models with your brand assets

Train custom LoRA models with your existing brand assets to generate visuals that closely match your established brand style.

📍 How to Access

Navigate to Visual Foundation → Train or go to /visual-identity/train.

💡 Training is Optional

In the Visual Foundation workshop, style selectors (icon, illustration, pattern, imagery) already guide brand-aware generation when creating—no training needed. Training is optional for stronger style matching. Logos require no training; save them to your asset library.


What is LoRA Training?

LoRA (Low-Rank Adaptation) is a technique that fine-tunes AI image models with your specific assets. This enables:

  • Style Matching: Generated images match your existing visual style
  • Consistency: All outputs maintain brand coherence
  • Efficiency: Fast training with relatively few images

Requirements

⚠️ Before You Start

You need at least 5 images per direction (imagery, illustration, pattern, or icon). 10–20 usually look better. Logos do not require LoRA training.

Image Requirements

  • Quantity: At least 5 per direction (10–20 recommended)
  • Quality: Clear examples of the style you want to repeat
  • Variety: Different subjects or compositions in the same style
  • Format: PNG or JPEG. SVG is allowed for icons, illustrations, and patterns — Eklipsa converts SVG to a 1024px PNG before training
  • Size: 1024px or larger looks best; smaller images are scaled up

Training Process

1
Upload & Configure Your Assets

Upload your brand images. The system analyzes them for training suitability.

2
Configure Training

Set training parameters or use recommended defaults. Options include training steps and learning rate.

3
Start Training

Begin the training process. This typically takes 15-30 minutes depending on your settings.

4
Review & Activate

Review sample outputs. If satisfied, activate the model for use in your visual tools.


Using Your Custom Model

Once trained, your custom model is automatically used in:

  • Icon Generator
  • Illustration Generator
  • Imagery Generator
  • Pattern Generator
  • Any visual generation tool that supports LoRA

Logos do not use LoRA training—they are generated with Ideogram Transparent and saved to your asset library.


Using Brand Style Mode

Once you've trained a custom model for imagery or illustrations, you can use Brand Style mode in the generators. This mode explicitly uses your trained model to generate visuals that match your brand's established style.

💡 What is Brand Style?

Brand Style is a special option in the Style dropdown of Imagery and Illustration generators. When selected, it uses your trained LoRA model instead of generic style presets, ensuring maximum brand consistency.

How to Use Brand Style

1
Train Your Model

Complete the training process for imagery or illustrations in Visual Foundation (Visual Foundation → Train). You'll need at least 5 training assets selected.

2
Open the Generator

Navigate to Halo Studio → Imagery Generator or Illustration Generator.

3
Select Brand Style

In the Style dropdown, select "Brand Style" (it appears at the top with a checkmark icon when your model is trained).

4
Generate

Enter your prompt and generate. The tool will use your trained model to create visuals that match your brand style.

📍 Visual Indicators

The Brand Style option shows different indicators based on your training status:

  • ✓ Checkmark: Model is trained and ready to use
  • ⚠️ Warning: Training assets selected but model not yet trained
  • 💡 Info: No training assets selected yet
💡 Pro Tip

Include variations of your assets (different colors, orientations, contexts) for a more versatile trained model. Train separate models for imagery and illustrations for best results.

Was this article helpful?