AI Training

AI Training

Training data for Image Generation with Stable Diffusion

AI Demo

Intermediate Images

My History with AI

I have been interested in modern AI (LLMs and GPTs) since late 2021 when I discovered a game called AI Dungeon. Not much later I discovered DALL-E, and Midjourney before later transitioning to Stable Diffusion. In the meantime, I became an early adopter of ChatGPT, which I still use regularly. Finally, I continued to play with image generation and began learning the basics of creating image models and Loras.

AI Dungeon

I have been following the development of AI since late 2021, when I discovered a game called AI Dungeon. I am a huge fan of tabletop games like Dungeons and Dragons, and was blown away by the ability of the AI. It was still primitive, but the closest thing I had seen up until that point was Cleverbot, which is far less impressive by comparison.

DALL-E, Midjourney, and Stable Diffusion

Not long after the infamous Avocado Armchair started making its rounds around the internet, I became enthralled with DALL-E, though it was quickly overshadowed by Midjourney, as I felt it had higher quality images.

I enjoyed my free trial of Midjourney, but once it ran out, I found myself wanting to self-host an image generator and found Automatic1111, which is a Web UI for Stable Diffusion. I set it up on my server and was able to, slowly, generate my own images.

I have a few of the images I generated back then in my AI Gallery if you're interested.

ChatGPT

Once ChatGPT went public, I got on board almost immediately, signing up within the first few days. As a programmer, I really like how it can help with complex topics and write tedious code snippets very quickly. I've been a subscriber since it launched, but I would like to self-host another model like Mixtral 8x7b, but its performance on my server is just too slow. I guess I need a stronger GPU than an Nvidia Tesla for this one.

Training the Model With My Own Face

I wanted to learn more about how AI models are generated, and took the opportunity to train a model using my own face. It was a fascinating process, and it gave me a much deeper understanding of how these technologies work.