Models with adjusted training parameters will automatically be classified as Custom. Adjusting parameters is for advanced users only, after having tested the other presets.
Scenario cannot guarantee results from custom training parameters, which are recommended only for advanced users.
Some advice for users who are considering adjusting their training parameters:
- Try small, incremental adjustments.
- It is very rare to require an extremely low or extremely high setting on any of the sliders
It is recommended to improve your captions and your dataset while still using the Scenario training presets before adjusting advanced training settings.
Basic Tips for Curating and Captioning a Dataset
Guide to Manually Captioning Your Dataset
Training Steps
The Training Steps refer to how many cycles the AI takes to go through all the training data. The Text Encoder Training Ratio represents what percentage of the total steps are spent on the text encoder.
Consider the Training Images in Your Dataset
Typically, it is wise to use fewer training steps with smaller image datasets, and more training steps with larger image datasets. The default (recommended) preset on Scenario is 175 training steps for every image.
Adjust Incrementally
The default formula on Scenario is 175 training steps for every image. This means if you upload 10 images, the recommended total steps will be 1,750.
For datasets larger than 30 images, you may usefully experiment with reducing the number of steps per image
Best Practices
If you are considering adjusting the training steps, here are some best practices:
-
Make small incremental adjustments.
-
When increasing the learning rate, decrease the training steps, and vice versa
Learning Rate
The Learning Rate has to do with how fast or slow the AI learns and contextualizes the training data (dataset images and captions.)
UNet and Text Encoder
In general the UNet controls the amount in which the images in your dataset are trained in your model and the Text Encoder takes into account the captions provided.
Best Practices
If you are considering adjusting the learning rate, here are some best practices:
-
Make small incremental adjustments.
-
When increasing the learning rate, decrease the training steps, and vice versa.
Text Encoder Training Ratio
The Text Encoder Training Ratio represents what percentage of the total Training Steps are used to compare the image captions with the training images.
What does the Text Encoder Training Ratio do?
The larger this number gets increases the amount the text encoder influences the overall training. This will in turn result in stronger caption influence.
Clarifying Concepts
The AI model already understands a lot of concepts; it understands a lot about what words relate to what images and actions in those images. Training the text encoder does not typically teach brand new concepts, instead it helps to make the relationship between what the AI knows, and what it does not recognize in your image more clear.
Styles are Lower
Because the AI already knows many concepts, it is typically advised to keep the text encoder training ratio lower for style training.
Subjects are Higher
Conversely, it is important to increase the ratio for subject training, where the concept being learned has very specific details which need to remain incredibly consistent.
The Text Encoder Training Ratio is considered an advanced parameter; the more you understand how AI models are trained, the more creative your can be with advanced parameters.