Training a Custom Subject Generator

We will guide you through one method of training a Custom Generator for Player Characters (Subject)

     In our Concept Art training guide, we went over three main categories your custom generator might fall into. We recommend you review that guide, but as a recap those categories are style, subject, and hybrid.

This player character guide will be focused on subject training. That means you should think of the custom generator you make from this as being a detailed and nuanced illustration of a very refined and well defined subject in your world. Character training has a small learning curve, although it can be an incredibly versatile tool. Feel free to read more about regularization classes here.

Creating Stargirl

     This tutorial is going to focus on training a player character custom generator using the original character, Stargirl. We’ve provided a link to our dataset and welcome you in following along in our training. Or feel free to follow along using your own original character.

     It is important for the sake of this subject guide that the final result be a character who is able to fit into any aesthetic style she is introduced to, without bringing in too much disruption. This is particularly important when you are training a larger model, or creating storyboarding. Style-free subjects, such as characters, can be used to create promotional graphics in various aesthetic mediums, and have a lot of potential for long-term, cross platform use as avatars.

Curating the Dataset

     When you take a look at the dataset provided, take a moment to notice - what is the similarity in each image, and also, what makes that image different from the rest of the dataset?

     When you look at the images, there should be a very notable and distinct set of similarities:

  • Height, hair color, eye color, and clothing never change.
  • Regardless of the change in aesthetic, all of the aspects that make this character uniquely them remain.

     What makes each image distinct and different is also important. What we notice here are the range and difference in the variety of styles and medium present in the dataset.

     As a general rule the things that are shared throughout a finetuning dataset are prioritized in the training, and the things that differ are not. This isn’t a perfect rule - there are exceptions - however it is a good starting point and rule of thumb.

     You will also notice that there are 25 images in this dataset. Although quantity is very important, particularly for style and concept training, it is equally important what you put in. This model could be trained on 12 images without issue - I encourage you to try it out! In this case, I am focused on getting more nuance, so I’ve opted for a larger and more diverse dataset.

     The minimum number of images you should use is five. However, we don’t recommend going below eight, as your images are more likely to underfit the fewer training steps you have. It will depend on what you are training, but somewhere between 10-30 tends to be ideal.

     It is important to remember that you may need to refine and retrain your dataset, particularly when you first start using the program.

Create Your Generator

     Once you have your dataset, you’ll be ready to create your generator. You will need to go through the following steps:

  • Go to Home > Create a Generator
  • Upload your images to Add Training Images. If you need to crop them, do so at this step
  • For this model we will remove all backgrounds using the Remove Background button
  • Click Next when you are ready
  • You will not be prompted to choose a name and a regularization class. In this case we will be naming the model Stargirl and picking Avatars/People > Person as our regularization class.
  • Click next. You will see the next page which talks about learning rate and text encoding. You do not need to change anything here, and you may launch the training program.

         It should take anywhere from 20 minutes to around 2 hours for your generator to train. Check back in a little bit!

Test Your Generator

     Once your training is complete, it’s time to test your outputs. Follow the guide we’ve created in our document on How to Prompt. If you are happy with your work, proceed to the next step. In some cases, you may find you want to adjust the outputs. You are always welcome to add or remove images and retry.

     If you are not happy with your output there are a few ways to proceed.

     First, you may decide to use an imperfect generator with some extra prompts to create more nuanced images for your dataset. You can use these images to replace or add to your original dataset when you train a new model.

     Your other option is to remove images from your original dataset that you see are showing up too often. It is easy to identify these if, when you generate images without additional prompts, you see one or two aesthetics or subjects from the dataset coming through multiple times. That indicates a need to prune and retrain.


     You can see that with a little practice, it can be easy to train a character. Try to remember - range in style is good! You can also use this guide and ignore that recommendation if you only plan to keep the character in one aesthetic. In that case, you would want to ensure the range is present in face expressions and poses. We can’t wait to see your results - make sure to tag @Scenario_gg on twitter so we can see what you’ve made!