khyle. LoRA Style

About this LoRA

This is my first style LoRA, trained specifically for the Illustrious XL ecosystem.
While it already delivers good and recognizable results, it is still an early version and will be improved in future iterations.
Some prompts may require slight adjustments to achieve the best output.

Recommended Usage

- Trigger word: Use '(khyle)' — preferably at the beginning of the positive prompt for stronger effect.

- Best starting prompt:

(khyle), masterpiece, best quality, traditional media, (anime coloring, anime screencap:1.5),

- LoRA Strength: I used 1.0 for all the example images on the model page. However, using a lower strength between 0.4 and 0.7 can also be a good compromise depending on your prompt and desired intensity. Feel free to experiment to find what works best for your generation.

Training Details

- Base model: Illustrious XL 0.1
- Training tool: ai-toolkit
- Resolution: 1024px
- Epochs: 10
- Network Dim / Alpha: 32 / 16
- Optimizer: Adafactor
- Learning Rate: 0.0005 (Text Encoder: 0.00005)
- Scheduler: Cosine with min SNR Gamma 5
- Other settings: 6 repeats, token shuffling, flip augmentation, noise offset 0.1, trained with TextEncoder

All example images in the gallery were generated using WAI illustrious SDXL model.

Note

This LoRA is designed to work best within the Illustrious XL family.

Future versions will aim for better prompt adherence, stronger style fidelity, and more flexbility.

Feel free to experiment, but starting with the recommended trigger and strength will give you the most consistent and satisfying results.