lydrasdxl - SDXL LoRA

Trigger word: lydra

Overview

This is a high-quality SDXL LoRA trained on a curated dataset of 25 images featuring a distinctive character with consistent facial features and versatile poses.

The dataset includes a balanced variety of angles and expressions: front view, right profile, left profile, and three-quarter views, ensuring the LoRA performs well across different compositions and lighting conditions.

Recommended Usage

- Trigger word: lydra

- Sampler: DPM++ 2M Karras

- Steps: 25-35

- CFG Scale: 7.0

- Resolution: 1024x1024 or 896x1152 (for portraits)

- Hires Fix: Recommended (1.5x-2x with 4x-UltraSharp, denoising 0.4-0.5)

Training Details

- Base model: SDXL 1.0

- Dataset size: 25 images

- Training steps: 2000

- Checkpoints: 10 (saved every 200 steps)

- Resolution: 1024x1024

- Learning rate: 1e-4

- Optimizer: AdamW8bit

- Batch size: 1

- Mixed precision: fp16

Disclaimer

IMPORTANT: This LoRA depicts a fictional character created for artistic and creative purposes. Any resemblance to real persons, living or dead, is purely coincidental. The character is not based on any real individual. The LoRA was trained solely on synthetic or fictional reference images.