Meet Lyssandra, an original female character

This is my very first LoRA, created as a learning experiment while discovering Kohya_ss and SDXL training. So please be a little forgiving if the results aren't perfect! 😅

I'm still learning, and feedback, advice, and constructive criticism are always welcome.

Training Information

• Base checkpoint: fubuwaMIX3DREAL v1.34 (SDXL)

• Training software: Kohya_ss

• Dataset: 17 images

• Training resolution: 1024px (with bucketing)

• Training epochs: 10

• Total training steps: 1,700

• Network Dim / Alpha: 32 / 16

• Optimizer: AdamW8bit

• Learning rate: 0.0001

• Training hardware: NVIDIA RTX 4060 Ti 16GB

• Training time: Approximately 34 minutes

How to Use

Trigger word: lyssandra

Recommended LoRA weight: 0.8–1.0

Example prompt:

<lora:Lyssandra_SDXL_v1:0.8>, lyssandra, 1woman, long reddish pink hair, amber eyes, detailed face, beautiful eyes, semi-realistic, cinematic lighting

What to Expect

The LoRA was trained on a relatively small dataset, so character consistency may vary depending on the checkpoint, prompt, pose, and other LoRAs being used.

My initial tests have produced encouraging results, particularly for portraits and semi-realistic illustrations. However, it hasn't been extensively tested across different checkpoints and styles.

The showcase images were generated using additional styling and detail LoRAs, so results may differ from the examples.

Final Notes

This is a first experimental release, not a professionally optimized model.

I mainly created it to learn the training process, experiment with character consistency, and share the results with the community.

Thanks for checking it out, and have fun experimenting with Lyssandra! ❤️