Important: Use the tag momo_ayase to make sure the LoRA represents the character!
If you want the original style to be taken more into account, also use nux_momo.
This LoRA was trained with a focus on practical usability and variation. The dataset is compact but carefully curated, prioritizing quality of representation over sheer volume.
_______________________________________________________________________________________________
Training Setup:
Images: 47
Batch size: 6
Steps: between 1600–2200 (depending on version)
Optimization: Standard parameters with adjustments for stability and consistency
______________________________________________________________________________________________
Concept:
The goal was to create multiple versions that balance creative freedom and accurate representation. Some versions are more relaxed and versatile, while others are deliberately overfitted to deliver highly precise but less flexible results.
______________________________________________________________________________________________
Usage:
For creative, free-flowing scenes, use the loose version.
For general prompts, the balanced version is the best all-round choice.
The overfit variants are useful when maximum accuracy or a very tight style is required — but keep in mind they are less adaptable to new prompts.
This LoRA is intentionally released in several versions so users can choose between flexibility and precision depending on their needs