📌 SDXL LoRA – Custom Client Request (by Dead_Sec)

This is the SDXL version of a custom LoRA trained by Dead_Sec for a private client request and now shared for broader use.

The LoRA was trained on a carefully curated dataset of up to 30 images, with the dataset generation and refinement assisted by ChatGPT to ensure consistency in concept, composition, and visual identity. Special care was taken to maintain clean subject definition and stable results across different prompts.

🧠 Training Details

• Base Model: SDXL

• Training Tool: kohya_ss

• Epochs: 10

• Batch Size: (default kohya_ss configuration)

• Learning Rate: 0.0001

• Dataset Size: ~30 images

The training setup was optimized to balance concept retention and flexibility, making the LoRA responsive to prompt variations without overfitting.

🎨 Usage Notes

• Designed for SDXL checkpoints

• Works best at LoRA strength between 0.6 – 1.0 (adjust based on style intensity)

• Compatible with both photo-realistic and stylized outputs, depending on prompt structure

⚠️ Disclaimer

This model was originally trained for a client request. While it performs well for its intended concept, results may vary depending on the checkpoint, sampler, and prompt style used.

If you enjoy the LoRA or find it useful, feel free to leave feedback or example generations—support helps improve future releases 🙌