Description
The Oracles
The Oracles is a generative AI PFP collection. The project brings together some of my creative explorations over the past 25 years, in portrait painting, creative coding, automatic drawing, generative NFT’s and AI. The collection began through playful experimentation with AI platforms Midjourney, DALL·E 2 and DreamStudio that utilize a newer AI approach generally termed diffusion. The diffusion model works by progressively adding noise to an image and then utilizing a neural network to denoise. There is also a dataset of billions of text-image word pairs the system is trained on. Iteratively applying the noising/denoising process in concert with the text-image training enables diffusion to synthesize completely original images.
As with most of my creative practice, I had no fixed plan or project goal when I set out playing with diffusion, just pure curiosity to explore this new powerful approach. You interact with the diffusion platforms by inputting text prompts with additional weighting values. They can also optionally take image input. This image option quickly became fundamental in my exploration, and I began feeding my drawings, paintings, portraiture, and creative coding output into the different AI platforms.
My early results were startling. I had been aware of some art created with earlier GAN AI image systems, but I hadn’t made a connection to my own creative practice. The GAN work I saw seemed to have a strong AI aesthetic, which felt too imposing for my interests. Using diffusion, I was able to generate very subtle variations of my existing pieces that felt authentic, without sensing a too dominant AI aesthetic. (Of course, diffusion does too have its own stylistic biases and limitations.) Subtly playing with the combination of input image, text prompt and weighting, I discovered that not only could I create strong individual pieces, but a cohesive series with rich variation.
As a painter, I draw great inspiration from work of the early Renaissance, especially painters Piero della Francesca, Andrea Mantegna, and Giovanni Bellini. My very earliest experiments with diffusion involved inputting my figurative drawings and including the names of these three painters in the text prompt. I began seeing subtle variations to my drawings that felt as if they (and I) had been transported back 600 years. This was unexpectedly deeply moving to me personally and triggered a much deeper engagement with diffusion.
The experimentation eventually led (back) to a focus on portraiture. (Though I still had no intention of creating a PFP collection.) The more I played, the more a larger series of heads began emerging. The input images included my drawings, paintings, and photos, most often of family members. The wide range of outputs from this experimentation included the truly horrific and photorealistic, to much softer, poetic imagery on the other extreme. I found myself seeking some middle aesthetic ground, informed by the early Renaissance.
Once I had the seeds of a new series, I began to consider how to produce it and ultimately how to release it. There were a number of technical hurdles that needed to be addressed. The web-based platforms mentioned earlier (Midjourney, DALL·E 2, DreamStudio) are not free services, but require the purchase of credits. Though the cost is very reasonable, the credit structure didn't encourage a more expansive experimentation, especially along aesthetic edges, where much of the output was undesirable. Some of the platforms have usage restrictions for the work created using them, as well as NSFW filters, which block a percentage of outputs. There are also serious limitations around sheer compute power using the online services, and diffusion in general is not a fast process.
One of the large groups organizing the diffusion promotion and dissemination is stability.ai, who produce a diffusion implementation called ‘Stable Diffusion’. Fortunately, Stability.ai decided to open-source their diffusion implementation, which meant I was able to run it locally on my own machine, but even more importantly, I was able to install it on my university high performance computing cluster. This allowed me to run diffusion for free, with no censorship and in large batches of images. Additionally, as a coder, I was able to write my own scripts to efficiently design an Oracles development pipeline.
The last technical challenge, which was a sizable one, was how to release a generative AI NFT with no existing platform. Fortunately, Emergent Properties (‘EmProps’) came to my rescue. I had worked with the EmProps team on an earlier generative NFT project, and we had developed a very good working relationship. The timeline for my project also fortuitously aligned with some of their platform goals. They agreed to build the very first generative AI NFT platform on the Tezos blockchain. The platform enabled me to utilize JavaScript to create a program that randomized image input, text prompt creation and weightings, which then gets automatically fed into Stable Diffusion behind the scenes. So essentially, they’ve built a modular, super generative NFT pipeline, the very first of its kind.
I am immensely grateful to the EmProps team, SMU Center for Research Computing and Stability.ai for enabling me to release Oracles. Thank you also to my dear family for putting up with the even more insane hours than usual.