LyCORIS-experiments
| Property | Value |
|---|---|
| License | CreativeML OpenRAIL-M |
| Author | alea31415 |
| Default Settings | LoHA net dim 8, conv dim 4, alpha 1 |
What is LyCORIS-experiments?
LyCORIS-experiments is a comprehensive investigation into different training configurations for character and style transfer learning in stable diffusion models. The project explores various architectures including LoRA, LoHA, and LoCon, with detailed analysis of hyperparameters, base models, and training methodologies.
Implementation Details
The experiments use a default configuration of LoHA with network dimension 8, convolution dimension 4, and alpha 1. Training employs a constant learning rate of 2e-4 with Adam8bit optimizer at 512 resolution and clip skip 1.
- Multiple character training examples including Anisphia, Euphyllia, Tilty, and OyamaMahiro/Mihari
- Extensive style transfer experiments across different base models
- Comparative analysis of LoRA, LoHA, and LoCon architectures
- Investigation of training resolution, learning rates, and optimizer effects
Core Capabilities
- Character fine-tuning with style preservation
- Style transfer across different base models
- Base model compatibility analysis
- Optimization strategy evaluation
Frequently Asked Questions
Q: What makes this model unique?
This experiment provides comprehensive insights into the behavior of different LoRA variants and training configurations, with detailed analysis of model transfer capabilities and style preservation across different base models.
Q: What are the recommended use cases?
The model is particularly useful for researchers and practitioners looking to understand optimal training configurations for character and style transfer learning. It provides valuable insights for choosing base models and training parameters based on specific use cases.
