Messing with Machine Learning image generation started in January, 2021. I started with existing StyleGan models on Runway, and soon started training my own.
VQGAN + CLIP came shortly after. Generating video from stills using text prompts felt like wizardry. Something special had unleashed on the world of creators.
Here are some quick showcases of what I have explored:
Now for the new stuff.
I received my invite for the private Midjourney beta in June, 2022. The tools is intriguing to use. It operates as a Discord bot that users can interact with through submitted prompts. After submitting a prompt, the bot returns a set of four images within roughly 60 seconds.
From these four images you can choose to re-roll with a new seed, create a variation of 1, or upscale 1. After an initial upscale you can choose to further "max upscale" the output.
And you know what?
The results are amazing.
Midjourney is the first off-the-shelf tool I have used in a long time that embodies Arthur C. Clarke's often quoted turn of phrase:
Somewhere along the line I started testing ML image output using various quotes from Lord of the Rings, and The Hobbit. For these, I used the lines of the first stanza of
Far Over the Misty Mountains Cold.
I was blown away when I saw these. And then I wondered what would happen if I used the entire stanza as a single prompt:
These developed with such a sense of dynamism. It was an exstatic feeling watching Midjourney materialize these.
But Midjourney can do far more than create an illustration from a book excerpt. It can also borrow from the style and form of anything in its training model.
As an initial forray, I decided to generate some "tintype" photographs.
Simply amazing. I am so excited to have gotten access to the Midjourney beta. Time to imagine.