Embedded Files
In the summer of 2021, a google colab notebook came into my hands, which using VQGAN + Clip (paper), was capable of generating images from a text sentence, the results were quite impressive, and there was something magical in writing a sentence and see how the AI interpreted it and generated what for me was undoubtedly a work of art. I was hooked. I kept testing for months, using initial images that the AI used as a starting point, mixing text inputs so that the image was a combination of two sentences, and experimenting with a new concept for me, prompt engineering, which is to create sentences from input in a certain way, so that they generate more interesting results, or apply a different style to the generated image. Some typical examples can be ending the sentence with "top of artstation" to give it a professional drawing style or "rendered by Unreal Engine 4" so that the image looks like it was generated by that engine. Some of these images can be seen in the following instagram account, where I uploaded some of my first tests, using titles of books, movies, comics and video games as prompts.
At one point I started using as a starting image a sketch I made some time ago of Jax, the little robot that appears in Ted Chiang's "The Life Cycle of Software Objects", a book I recommend to everyone, especially if you are interested in Artificial Intelligence and you like science fiction and also to include "Rusty Roboz" in all the prompts. The idea was to see if from the same initial image and the words "Rusty Roboz" interesting variations could be created. In this way I started to create a lot of Rusty Robozs, which took a long time since running these models requires a good graphics card, which I did not have, so I had to settle for using Google Colab, an excellent service that provides free access to a notebook type development environment with access to graphics cards for free, has some time constraints, and the graphics cards they offer are not the most powerful but it's a great way to experiment with this kind of algorithms free form. Some I let them generate for only a few hundred iterations and others I left them for hours, reaching several thousand iterations, I was trying variations of the same prompt, colors, objects, etc. Little by little you start to get a kind of intuition about which prompt will generate the most interesting output.
Original sketch of Jax
Here are some generated images and some videos:
From text: Rusty roboz with a ligthsaber in a volcano
From text: Red roboz surfing in the sea
Ciborg with neons
Luke Skywalker
Blue rusty roboz space pirate
Dune
Assassins Creed
Uncharted 2
GTA VI
batman dark knight returns
Superman red son
Brave little rusty roboz (500 iterations)
Brave little rusty roboz (2000 iterations)
Brave little rusty roboz (4000 iterations)
Brave red roboz on a hill
Brave little rusty roboz
After a while I had a considerable number of images and videos of different rusty robozs and inevitably it occurred to me to make a collection of NFTs.
BONUS: Some other generative models that I tried.
Google Sites
Report abuse