LyCORIS Techniques
An advanced alternative to LoRA that offers different fine-tuning methods like LoCon, LoHa, and LoKr. Can capture more complex styles and subjects with various precision/size tradeoffs.
LyCORIS (Lora beYond Conventional methods, Other Rank adaptation Implementations for Stable diffusion) extends LoRA with additional fine-tuning approaches.
Types of LyCORIS
LoCon (LoRA for Convolution)
- Applies LoRA to convolutional layers
- Better for styles and textures
- Slightly larger than standard LoRA
LoHa (LoRA with Hadamard Product)
- Uses Hadamard product representation
- Better expressiveness
- Good balance of size and capability
LoKr (LoRA with Kronecker Product)
- Kronecker product decomposition
- More parameters
- Higher fidelity for complex subjects
Comparison
| Type | Size | Fidelity | Best For |
|---|---|---|---|
| LoRA | Small | Good | General use |
| LoCon | Medium | Better | Styles |
| LoHa | Medium | Better | Balanced |
| LoKr | Large | Best | Complex subjects |
Usage
Same as LoRA in most UIs:
- Place in LoRA folder
- Reference in prompt
- Adjust weight as needed
Some UIs require LyCORIS extension.
When to Use LyCORIS
- Standard LoRA not capturing style
- Complex textures or patterns
- Higher fidelity requirements
- Willing to trade file size for quality
Training
Many LoRA trainers support LyCORIS:
- Kohya’s trainer
- LoRA Easy Training Scripts
- Choose algorithm during setup