Discover more from Addition
Training Stable Diffusion On Myself
A couple of weeks ago, someone asked me for a headshot for a speaking engagement. I didn't have one – at least not one that was recent – so I considered my options. Could I use AI to generate my new headshot?
AI image generation models like DALLE-2 and Stable Diffusion have consumed an incredible amount of images from the web. But unless you're internet famous, these AI models will struggle to generate accurate pictures of you.
To get around this, researchers have learned how to train image models on new visual concepts. This technique is generally referred to as "Dream Booth." It was first published in a research paper from Google in August and has since been replicated by open-source developers using Stable Diffusion.
Training the Model
To get started we provide the model with a few dozen images of me, which we teach it to associate with a random keyword.
Once trained, the model can generate new images of me whenever we prompt it with my unique keyword. I can combine this keyword with a wide range of prompts enabling me to explore various styles.
Future Use Cases
Beyond solving my headshot problem, a wide range of commercial use cases will benefit from this technology as it continues to improve. For example:
Automotive companies that spend millions on 3D Renders may find more flexibility and creative possibilities training AI models to understand their new lineup of cars.
Advertisers could generate personalized ads that show me products in the situations most likely to appeal to me.
Future versions of this technology could be used to generate photo-realistic avatars of me within virtual environments like gaming and VR.
As always, reply to this email if you have any comments or questions.
Are you AI Curious?