Autopoeisic Transmogrification Fragments (2022) is a collection of short, seamless loops that blur the boundaries between biology, chemistry, and geology, an exploration inspired by reflections on the essence of life, intelligence, and consciousness, and on the intricate networks of interdependence across manifold scales of time and space that allow us to exist.
The work is a direct continuation of (and both conceptually and technically very related to) Distributed Consciousness. I wrote about the conceptual motivations behind this work in depth here.
(My interest in these themes continue in more recent work such as Boundaries and Superradiance, though technically these works use a very different workflow).
I had made the images in Distributed Consciousness using a custom VQGAN+CLIP workflow (this was before Disco Diffusion etc. And of course before Midjourney, ChatGPT etc.). This was also the same workflow I used for All Watched Over by Machines of Loving Grace (2021), and Deeper Meditations (2021).
As part of Distributed Consciousness, I finessed and released as NFTs, 256 images created this way. I continued to create more images in the same fashion, using the custom VQGAN+CLIP workflow I’d developed, and then I trained another GAN (specifically StyleGAN3) on these images.
These types of older GANs (like StyleGAN) allow smooth explorations of Latent Space (something which VQGAN, and more recent denoising diffuser based models cannot do very well).
This is something I had explored quite in depth in earlier work such as Deep Mediations (2018), and I wrote about in a blog post, and a paper published at NeurIPS.
Something else I had explored quite a bit – both in earlier work such as Deep Meditations, and in Autopoeisic Transmogrification, is to change the training data during training. I.e. start training on one dataset, and then swap out and switch to another training set for only a few iterations. This allows for some very interesting hybrid images, for example which retain large-scale structural or compositional elements from one dataset, and finer detail from another.
Having trained this new GAN model on the still images I had created, I applied latent space explorations and created these smoothly morphing, evolving autopoeisic transmogrification loops where the boundaries between biology, chemistry, and geology are blurred.
The sound design in all the pieces was made the old-fashioned way — just me, working manually with a bunch of synths.
Fragment #000 was minted Oct 4, 2022 for an exhbition curated by GANBrood and Georg Bak for Kate Vass Galerie in Zurich called ‘Freak Show‘.
Fragments #001-#005 were minted Jul 24 – Aug 17, 2022 for the launch of the Versum platform.
They can all be found onchain at https://objkt.com/tokens?search=autopoeisic%20transmogrification