This is algorithmic, generative sci-fi pulp from the 70s, courtesy of OpenAI. Artist Lewis tells us how it was done. In the meantime, see how many friends will pretend to have heard of Neytiri A. Quaritch’s pioneering tome, Green glass is the color of the wind.
Whether you’re a die-hard sci-fi fan or just rummaged through 70s paperbacks at a flea market, I’ll give you a moment to soak that up before getting into the techniques explanation. machine learning behind the scenes. I had to clean some coffee which I chuckled.
The way a lot of press gets it wrong, of course, is to say things like “AI made science fiction book covers.” Even though these algorithms are getting much more sophisticated than pixel averaging or a Markov chain, they are still just algorithms, lacking agency, albeit with huge data sets as their source material. In turn, however, this makes some of the aesthetic peculiarities they generate all the more interesting, and means that it is useful to understand them as generative tools in the hands of artists. They’re the result of a lot of human effort in math, code, and ultimately human choice, though that last bit bothers those looking for general artificial intelligence.
Lewis Hackett is that artist and has cleverly curated what we see, combining a graphics technique called Clip Guided Diffusion for imagery with familiar GPT3 techniques for titles. And he did a great job selecting the results and aping the typographic style by hand.
Lewis explains his process to the CDM; I’ve added links to the relevant machine learning notebooks and code:
I experimented with various Google Colab notebooks to see how I can integrate AI/GAN [Generative Adverserial Networks] techniques in my creative workflows, both for generating ideas and finding interesting ways to use them conceptually.
While exploring CLIP Guided Diffusion Notebooks, I came across ‘nshepperd’s JAX CLIP Guided Diffusion v2.3’ which is based on Katherine Crowson’s CLIP Guided Diffusion Notebook. I’m a huge fan of 70s sci-fi artwork and have tried various prompts around the theme of “70s sci-fi book cover artwork” which have already produced amazing results. I refined the prompt and ran a batch of about 100 variations with different starting values.
Then I started testing OpenAI’s GPT-3 since it became publicly available in November 2021. I opened a chat conversation and asked him for “title ideas for my science novel -fiction”, and I followed other queries such as “something more”. mysterious” or “something less obvious” until he started giving me interesting titles. I repeated the same process, but instead asked him for fake author names.
Once I had the titles and authors, I used them to help organize the final collection of generated artwork and created the typography using popular fonts from the 70s. also used Waifu2X to scale the artwork as it was originally generated in low resolution. I think the final releases are fascinating because they convey that same sense of mystery that book cover artists use to get people to pick them up and read them.
I’m also interested in the concept of having a complete product generated with AI techniques. While this required my input and curation, I expect this to diminish over time as technologies advance and we finally see the electric sheep that androids dream of,
I tend to be allergic to the idea that technology will advance, but there’s concrete reason to believe he might be right about AI-generated work – partly because you might train the ‘IA to self-direct some of the conservation he describes. You may still need to give an end goal, but the process could theoretically involve at least fewer interventions. It’s an intriguing question.
Meanwhile, these covers are just… hilarious. And given that we’ve had a few generations of tools giving us pretty obvious choices in tool sets, they also exemplify a world in which creative art tools could be a lot…well, more trippy than they are not now.*
On the subject of GPT-3, here are some essential reads from AI journalist Will Douglas Heaven of MIT Technology Review:
The new version of GPT-3 behaves much better (and should be less toxic)
More – enjoy!
To follow Lewis on Twitter.
Seen previously doing some nice things with Unity.
I wonder what a Kai’s power tools would look like powered by this kind of AI techniques instead of the fractal and channel operation techniques that Kai Krause used to create branching possibility trees. Kids, ask your parents or Wikipedia. Speaking of blasting the 90s… KPT is now in PaintShop Pro, sold by Corel, in a new version. Paintshop Pro? Kai? Corel? Now, there’s a name I don’t have… Oh yeah, and it has an AI too, so there you go.