In this article we will walk through a simple workflow for training your custom generator without a regularization class, with the addition of text encoding
This guide is intended to be a framework, and may need to be changed slightly to accommodate individual datasets.
Typically when training models, users find themselves indicating a regularization class. You can see examples of a guided training with a regularization class in our concept art training example. However, it is entirely possibly to train a generator without a regularization class. Typically we only recommend these workflows for users who feel confident in their ability to train basic models.
A regularization class improves the accuracy of neural networks by introducing noise or randomness to a set of data points, such as "pixel art" or "vector images". This helps to prevent overfitting, where a neural network becomes too precise with a specific training set but is unable to make accurate predictions on new data. Using regularization images during training can help a neural network to detect more general patterns and improve its overall accuracy.
With the adoption of text encoding it is possible to still get a good output from training, even without a defined class. When using a text encoder it's important to use fewer steps and adjust your learning rate. It is far easier to overfit a model when text encoding is active, in part because the GenAi training recognizes and infers more quickly within those training parameters.
Recommended Workflow
Advanced users may find a variety of different workflows suit their needs depending on their desired output. In this guide we will present a standard workflow which should be used as a baseline. Following the steps for most general models of** 20 or fewer images** should result in a positive output. We recommend only adjusting the parameters by small step or learning rate changes to adjust.
Training Parameters
Training Class: Unguided
Training Steps: 100
Learning Rate: 5e-6
Text Encoder: On
If you increase your dataset beyond 20 images, we still recommend starting with this training parameter to get an idea of what sort of output you will see.
Walk Through
Step 1 - Upload Images
Navigate to your generator creation screen. Upload five or more images - in this example you will see we are using only 10 images in our dataset. Click next.
Download Our Dataset
Step 2 - Mode Selection
On the next screen, chose "Advanced Mode" and click next.
Step 3 - Name Your Token
On the next page pick the Unguided category and a name for your token. If you choose a token name that is easily relatable to a subject, in this case astronaut, it may cause unexpected results. We recommend naming your token something that the training data likely won't have a relation to.
Step 4 - Set Up Your Training Parameters
Your last step should be defining your training parameters, which I shared above. Make sure the text encoder is on!
You can then start your training process. We recommend using our prompt guide to test the functionality of your trained model, and making small adjustments of 20-50 steps, or increasing or decreasing your learning rate to get your desired outcome.
This workflow is better for generators that are meant to capture a style rather than unique subjects, and tends to give better results when used for more general models.