Skip to main content
All CollectionsHelp Guides
Creating Consistent Characters
Creating Consistent Characters

Image Guidance, Character Reference and Training a Custom Element

Nick Lee avatar
Written by Nick Lee
Updated over a month ago

Maintaining uniform AI-generated character designs is useful in a variety of contexts, from creating a cover for a book series, to generating concept art. However, it can be tricky to do – particularly when using an illustrative style.

To make the process easier, we have created this handy guide with everything you need to get you creating consistent characters. The guide features two different methods: using Image Guidance and Character Reference, and using Model Training.

Method 1. Image Guidance and Character Reference

Step 1: Upload a reference image

Navigate to the Image Generation tool from the homepage. Click the 🖼️ button next to the prompt box, select Character Reference, and upload a well-lit, plain background face shot (real or generated) or choose an image from your Personal Feed. Click Confirm.

Step 2: Adjust character reference strength

Click the reference image (found at the bottom of the prompt box) to set its strength to Low, Mid, or High. Start with 'High' and adjust as needed.

Step 3: Create a detailed prompt

In the prompt box, describe your character's attributes and context. Example: "A male with short black hair, expressive eyes, large spectacles, wearing a baseball cap, standing in a city street." Detailed descriptions ensure consistency across generations.

Step 4: Add a Preset and Elements

Select your preferred Preset Style and include Elements (optional) or any other Image Guidance options such as Style Reference, then hit the generate button to generate your image.

Step 5: Refine and evaluate

Review images for precision and consistency. Adjust settings as needed. Use Quality mode for better images and reduce strength for further stylization if using a photorealistic face.

Tips:

  • Character Reference is most effective with photorealistic faces or characters generated from the same model.

  • Ensure you have legal permission to use the images.


Method 2. Training a Custom Element (LoRA)

This second method requires more time and effort, as well as a set of source images, but can yield extremely effective and consistent results when done correctly.

Because training a custom Element can be an in depth process, this guide will, focus more on the specifics of Element training necessary to achieve a consistent character. Click here for a more detailed guide on Element Training.

​Step 1: Start Training

Navigate to the Models & Training section, then click on Train New Model.

In the Select Category menu, select the Character category, then click Next to continue.


Step 2: Upload Reference Images


The modal should take you to the Create Dataset section. Else click on Create New on the top left to continue. Upload at least 8-10 reference images to help generate your baseline character. These images should represent the characteristics of the character you wish to create.


It is recommended to keep most images to mainly closeups of the face, with a few full body and medium shots to help with additional information and if there is a good variety of lighting and color treatment / styles.

Good dataset: An example of a character dataset that has a good mix of lighting and shot types.

Bad dataset: An example of a character dataset with some poorly cropped images and similar images.


With character in stylized costumes (especially stylized characters), it is best to ensure there is more full body images so that the Element gets trained on it.

More information on datasets for character training for Elements can be found in the Element Training guide.

Note: You may use still stock footage, drawings, or photographs, but it’s important that you always have the legal permission to do so. (Whether that’s the license to certain images or permission to use a real person’s likeness).


Step 3: Train Your Element

After uploading the selected images, click 'Train Model' to begin the training process.
You will be presented with an popup that will guide you through the process.


In the Select Category menu, select the Character category, then click Next to continue. (This step is critical)

Select your character's dataset then click Next.


Enter the Element's name, description and trigger word. The trigger word is the most important and should be an abstract word such as jehgh. The click Start Training.

Note: Users on free plans are limited to training one Element at a time


Step 4: Generate Images

Navigate to the Image Creation tool, then under the prompt bar, select Classic Mode.


Select the appropriate compatible preset for your Element if it is not selected, this should be Flux Dev by default. Then on the left of the prompt bar, click on the picture icon and under Platform Elements, click on View More.

In the Elements popup, click on the Your Elements tab, then select your character's Element. Click on Confirm.

Once done you may start generating images of your character. It is best to be as descriptive of your character's appearance as possible for best accuracy when prompting.

The Element strength can be adjusted, however it is best to keep it at an higher value as lower strengths will result in greater deviations from the character's original appearance.

It may be likely that some images still won’t look exactly like what you want. Experiment with your prompt structure and description as necessary to refine and vary your results while maintaining consistency.

Step 5: Refine and Evaluate

Review your generated images for precision and consistency. Note that in a generation of 4 images, typically, 1-2 images may closely match the desired output, and the rest will not. That’s just part of the creative process!

Do not forget that you can try out different style presets or combine Image Guidance options and other Elements while using your very own character Element for even more creative control.

​​

Download successful creations and add them to your dataset to continually improve character coherence.

​Final Tips:

​Remember that when training your model for character consistency, the quantity and quality of your reference images directly influence the model’s performance.

Experimenting and tweaking your prompts and settings is essential to achieve the most best outcome. You’ll find that once the model has been sufficiently trained, producing a stylistically consistent character will get faster and faster!

That does it for our guide to creating a consistent character – we hope it helped! Remember, we're always adding new features and updating old ones, so be sure to check back in for new information and tutorials.

Happy Prompting! 🎨

Did this answer your question?