Dark Frontier: An AI Short

This started as an experiment with a LoRA I trained using a photoshoot with local model Katie, generating sci-fi-inspired images of her. One scene led to another, then sound effects, and after about a days work, this short was born. I could have refined the keyframes for better consistency in her suit and the ship, but this was a fun test for some bigger AI projects I have in the works!

A LoRA (Low-Rank Adaptation) is a technique used to fine-tune AI models, allowing them to learn new styles or subjects efficiently without retraining the entire model. It’s commonly used in AI-generated art to adapt existing models to specific characters, aesthetics, or themes.

To create the LoRA, I started with a high-quality dataset from the photoshoot with Katie, selecting 40–60 images—half close-ups to capture facial details, and the rest showing different angles and full-body shots for proportions. I trained the AI using Stable Diffusion on my local machine, adjusting key settings to balance accuracy and flexibility. The process takes about five hours on my RTX 4060 Ti, after which I get a model that can generate lifelike images of the subject in any setting. Sometimes, fine-tuning is needed, but this method has given me great results so far.

Here is another short video I created showing this process: