Creating professional headshots for Charlonis.com using AI has been an exciting mix of creativity and technology. While initial attempts using AITubo produced interesting results, they also revealed limitations. AITubo, for instance, allows generating LoRAs (Low-Rank Adaptation models) from a single image, which can lead to less realistic or versatile results. Expanding to more advanced tools, like Replicate, opens up the potential for building richer, more accurate LoRAs by training on multiple images—something I’m actively exploring.
What Is a LoRA?
A LoRA (Low-Rank Adaptation) is a model fine-tuning technique that adapts an existing AI model to capture a specific style, subject, or individual. Instead of retraining an entire AI model, a LoRA focuses on smaller, targeted adjustments, making it faster and more efficient. This makes LoRAs ideal for tasks like creating custom, hyper-realistic headshots that reflect personality while maintaining high-quality detail.
To build a LoRA that produces consistent and reliable results, it’s recommended to train it on at least 20 images. A diverse dataset ensures the model captures different angles, expressions, and lighting, resulting in a more robust and adaptable output.
Experimenting with LoRAs
My first experience with LoRAs began with AITubo, which allowed experimenting with headshot generation using just one reference image. While this was a good starting point, the limitations became clear—working with a single image made it difficult to achieve consistent or natural-looking results.
One-Image LoRA
I started with a recent pic taken on an adventure with some friends, who I cropped out so I had a single image of my face alone. In hindsight, I wonder if a tighter crop would have changed the results.
First Results
I enjoyed the snow scene and outfits AI decided to give me, but the first set of images are a whole lotta creepy, and WHY does AI think I’m going bald? Obviously, better results would come from training AI from a data set of more images of myself.
Sure takes a little extra courage to share self portraits that went weird...
Replicate, on the other hand, offers a more advanced approach. It allows for training LoRAs with a broader dataset, making it possible to fine-tune models for better detail and accuracy. I’m currently gathering selfies and other photos to create a LoRA that truly reflects my features and style, experimenting with various expressions, backgrounds, and lighting to build a diverse and effective dataset.
Stay tuned for better headshots!
What’s Next?
Expanding the toolbox to include advanced tools like Replicate, combined with the flexibility of LoRAs, marks an exciting step forward in crafting professional headshots. By working with a larger dataset, it’s possible to achieve results that are not only hyper-realistic but also highly personalized.
The process is a creative and technical learning experience, and each iteration brings new insights into how AI can elevate personal branding on Charlonis.com. With more experiments and refinements, the potential for achieving polished, professional headshots is limitless—and the journey to mastering LoRAs is just beginning.
