Training an AI model can feel overwhelming with all the jargon. Here’s a simplified explanation of the key terms in TensorArt training settings to help you get started confidently.
1. Repetitions
What It Means: The number of times an image is shown to the model during training.
Why It Matters: Repeating images helps the model "memorize" patterns and details.
2. Epochs
What It Means: One full cycle where the model goes through all the images in your dataset.
Why It Matters: More epochs mean the model has more chances to learn, but too many can cause overfitting (the model becomes too specific and less flexible).
3. Total Steps
What It Means: The total number of training rounds, calculated as:
(Number of Images) × (Repetitions) × (Epochs)
Why It Matters: More steps generally mean better results but require more time and resources.
4. Seed
What It Means: A number used to make random processes in the model consistent.
Why It Matters: Using the same seed ensures you get similar results when generating images.
5. Learning Rate
What It Means: How quickly the model learns during training.
Text Encoder Learning Rate: Controls how well the model learns tags or captions. Increase if the model isn’t recognizing key features.
Unet Learning Rate: Adjusts the model's overall learning speed. Be careful; a high rate can cause errors, while a low rate slows progress.
6. Grid Size
What It Means: The complexity of the model’s "thinking space."
Why It Matters: Larger grids allow more detailed learning but increase file size and training time.
7. Network Alpha
What It Means: Adjusts how much weight is given to changes in the model during training.
Why It Matters: Lower values make changes more noticeable, which can help refine details.
8. Scrambling Labels
What It Means: Randomizes the order of tags in your dataset.
Why It Matters: Helps the model avoid bias by seeing all tags equally.
9. Regularization
What It Means: Techniques to prevent the model from overfitting by limiting its focus on specific details.
Why It Matters: Regularized datasets help your model generalize better across different inputs.
10. Base Models
What It Means: Pre-trained models that act as a foundation for your training.
SD1.5 LoRA: Good for 2D characters.
SDXL LoRA: Great for realistic and detailed outputs.
Tip for Beginners: Start with the default base model (SD1.5 or SDXL) if you’re unsure.
Simplified Tips for Beginners
Start Simple: Use default settings at first to avoid confusion.
Experiment Slowly: Adjust one setting at a time to see its impact.
Review Regularly: Check your model’s progress during training and tweak as needed.
By understanding these terms, you’ll navigate TensorArt training like a pro in no time!
