Train a Character Model - For Beginners

Introduction

Custom models are a powerful tool for creating unique characters with endless variations and high consistency. While methods like IP Adapter or Midjourney “—cref” offer some consistency, custom models provide unmatched control over specific styles, subjects, or both.

Scenario makes creating custom consistent characters easier than ever.

Step 1 - Select Your Training Images

Curate a good training dataset.
‍
When training a character model, aim for a dataset of 10 to 15 high-quality images for optimal results. You can also use as few as 5 images or more than 20.

Step 1 - Select Your Training Images:

High-quality images: Use crisp, high-resolution images (minimum 1024px). Optionally, you can upscale your training images directly on Scenario to enhance their details. (Low-resolution training images will result in poor models)
Crop as square: All training images must be square. As you upload non-square training images to Scenario, you can adjust (crop) the part that will be retained for the training.
Consistency: Ensure your character remains "consistent enough" across images. The goal is for the AI to recognize the core characteristics of the subject, that are present in all or most images.
Diversity and variability: Keep enough variability in your training set, such as a variety of poses, zoom levels (portrait vs full-body…), expressions (happy, sad…), outfit, or backgrounds for example. Using too few or overly similar images will limit the model's variability and diversity.

Step 2 - Captions

Captions are labels that guide the AI in understanding your training images. While Scenario provides automated captions, it is highly recommended to manually review and adjust them to ensure consistency and quality.

Captions significantly influence your character model's behavior. It’s not necessary to caption the consistent elements you want to maintain (such as the hair color, or peculiar outfit details).Focus on captioning the variable elements like poses, expressions, or backgrounds.

Step 3 - Consider the style

For best results, use training images with a consistent style (realistic, cartoon, anime, etc.). This will help your character model generate outputs that match the style of your training images.

If you need to evolve the aesthetic style of your character later, you have two options:

Caption the style in your training images (e.g., add "cartoon" or "anime" in the caption).
Merge your character model with another style model, directly on Scenario (recommended method).

Step 4 - Hit "Train"

Keep the "Subject" presets. It's not recommended to use custom training parameters as a start - unless you want to train the model in a specific direction. Check out our tutorials on Training Steps, Learning Rate, and Unet architecture for more advanced customization options. [Link to tutorials]

Hit "Train" to begin the process. The training time varies depending on the number of training images provided and can take anywhere from 20 minutes to a few hours. Once complete, you'll have a custom character model ready to generate unique and consistent images.

Step 5 - Evaluate & Generate

Test your model with various prompts to ensure it captures the desired characteristics and diversity of your character. If you notice inconsistencies or areas for improvement, adjust your training data or captions and retrain the model.

Start with balanced prompts that resemble your captions. The easiest way to find relevant prompts is to use the "Prompt Spark" feature. You can start from a blank or just add a few words to guide the prompt completion.