Model Training: LoRa
Project information
- Category: Model Training
The aim of this project is to train a LoRA model for Yvonne Craig Batgirl from the 1966 Batman Series. I chose this character because it is very complex and my goal is to keep practicing with subjects that push the boundaries of the training even further. LoRA, or Low-Rank Adaptation, is a technique that allows efficient fine-tuning of diffusion models for specific concepts, such as characters, with smaller file sizes and faster training times.
I collected a diverse dataset of Batgirl, capturing the most varied range of poses and shots. I wanted to train a high-quality model, but it was a difficult project because the source images were blurry. I had to restore the faces one by one without losing the main features of Yvonne Craig. Some of her facial resemblance was lost in the restoration process, but the costume came out amazing. I think this is the best I can achieve given the quality of the source material. I created captions for each image, describing the pose, expression, and background, to serve as prompts for the LoRA model. I trained the model on Stable Diffusion XL, optimizing the settings and hyperparameters for quality.