As a curator and a collector, I must be able to assess the quality of an artwork. I decided to pursue artificial intelligence (AI) as a topic for this article because I didn’t feel I had enough information to appraise the medium successfully. I can’t code, and reading about neural networks makes my brain hurt. Personally, I like to imagine AI artists are like the Wizard of Oz, spinning wheels and turning knobs just out of view.
I am aware that there are resources out there, should I ever wish to educate myself on the science of machine learning, but for now, I am happy with my naive and somewhat theatrical understanding. Instead, I would rather focus my attention on the artists behind the emerald curtain. Thank you very much, Frank Poncelet, Ava Khan, Sqrt of Pi, NFTPhoenix, Cosmonutty and Rick Vito, for answering my many questions and furthering my comprehension of generative art and artists. From our conversations, I can now successfully describe AI art-making as a simplified four-step process: training, generation, curation, and post-production.
Training
All AI requires training. This is where an AI program is fed many different inputs, mostly images, in order to define parameters; for example, introducing a program to thousands of images of coffee mugs allows the AI to recognise what ‘coffee mug’ means visually. Some artists outsource this step by utilizing pre-trained neural networks found in some style transfer programs. Other artists use ‘secret programs’ to collect their immense datasets, which they then feed to their AI, and another utilizes ‘secret experts’ to train their AI for them.
“Having a well-trained neural network is paramount, and it is what enables you to have such control in tweaking the parameters.” -Rick Vito.
Outsourcing training seems common as it allows artists to focus on the other parts of the process where they feel more creative. Still, it is worth noting some AI artists feel training is a significant part of their creative process.
“In the beginning the artist is a teacher, compiling images for the model to learn from and adjusting the inputs as the outputs emerge. I collect my dataset by looking for public domain images similar to the category I want to generate. I download hundreds or thousands of them and then look through them 1 by 1.” -Sqrt of Pi
Generation
After training comes generation. Due to my limited technical knowledge, this is the part that remains most hazy to me. From my conversations, it seems artists can set their own parameters for things such as contrast and colour, but depending on how their AI was trained, they could also create sliders for things like less ‘coffee mug’ and more ‘German Shepherd’.
It’s in those thousands of small adjustments that the artwork is born, it is where the potential of the artwork is set free, and it’s where AI’s true power is revealed.” -Rick Vito
From how the artists talk about AI generation, it is clear that this step is where the magic happens, but who/what is the magician?
“During generation, the artist is literally piecing together a future artwork”-NFTPhoenix
Ava Khan describes generation more spontaneously.
“I picture the parameters but not the results usually. The results are my experiments, they are always unexpected, unplanned but amazing!”-Ava Kahn
But some artists described it more as a partnership.
“Many artists are interested in what the AI can do on its own, I’m trying to go the opposite way, and find out how well it can do when guided by us, and given enough love. I want a true collaboration.”-Rick Vito
Cosmonutty and was a bit more matter-of-fact.
“I play with the multiple settings to get a desired result I’m happy with.”- Cosmonutty
Sqrt of Pi views it as a personalised tool.
“AI is like the pan that someone has always cooked with and it carries its own flavor. My models have their own flavor that I’ve imparted over multiple failed and successful works.”-Sqrt of Pi
And Frank seems to enjoy the challenge.
“ Finding the right settings is the difficult part. There is not one solution that fits all.”- Frank Poncelet
I don’t usually feel like I am collaborating with my tools, so it seems strange to me that they refer to AI as a instrument and a collaborator. I have never painted a portrait and thought, wow, my brush really surprised me today! But, I suppose that this is just the evolution of ‘tool’, something useful that can now perform functions outside of our own narrow parameters.
“With generative art, I’m able to translate parameters that can’t be attained physically”-Ava Kahn
Curation
Just because your AI can produce it, it doesn’t mean it's pretty. Generation can output thousands of results. The artist must evaluate those generations and determine if there are any keepers in the bunch or if the parameters need more adjusting. Here is how the artists speak about curation.
“After you generate images with the model you assume the role of curator. You’re accepting or rejecting certain outputs depending on the style that you want the collection to have”-Sqrt of Pi
“Any generated artwork can be greatly improved by carefully tweaking parameters. The chance that a randomly generated image is the most interesting it could be, the most human, the most communicative, the most beautiful, dramatic or simply the very best it could be — the probability of that happening, without a human nudging the parameters for a few hours, is close to zero.-Rick Vito
“I try to push the machine to get the best possible iterations. I play with it until it pleases me the best.”-Ava Kahn
Generation and creation are iterative and it seems the artists toggle between them until they are satisfied with their results. For some artists, this is the last step in their process. Happy with their generations, they simply mint and list. Job done! For others, they choose to use some post-production flare to enhance their AI generations even further.
Post-production
Some adjustments add polish.
“I do paint over my results too, but I’m really trying to respect the output, so I only fix minor artifacts and repaint parts that didn’t do well in the upscaling.”-Rick Vito
Some adjustments are more extensive, adding depth and layering to the artwork.
“After generation, I touch it up again in photoshop and procreate to make a digital painting base that I prime for animation. Then I pixel loop animate the flow of the entire digital slate on the iPad and export into after effects to do more styling and morphing to get the trippy effects I’m happy with. The final step is taking the raw animation and colour grading it and adding sound design in the premiere.”-Cosmonutty
Videos seem to need more love post generation than still images.
“Finally if it’s a video piece, I render the model outputs along with my image edits, frame by frame, and stitch those together into a video.”-Sqrt of Pi
For video, the magic is all in the preprocessor. The AI will see the most slight changes in light between the frames, introducing a very strong flickering effect, people with epilepsy, would not be able to watch it. So the preprocessor flattens that to an acceptable level. -Frank Poncelet
After speaking with these artists I was amazed by how similar their processes seemed, and I was even more intrigued by how they varied. In each interview, I asked which step in this oversimplified process the artists felt afforded them the most creativity, and none had the same answer.
Cosmo believes he is most creative when conceptualizing before generation.
“Coming up with a story I want to tell or emotion I want to convey — what’s the soul of the piece and what do I want people to feel. That’s the most important step.”-Cosmonutty
NFTPhoenix feels most creative when assembling the training images.
“I feel every step has a piece of creativity, but especially when I think about the topic of my dataset, then pick good images and imagine the result, but a good result from this whole process directly depends on the creativity of the artist.”-NFTPhoenix
Sqrt of Pi likes the programming.
“I feel most like the artist when I’m writing the code to create the model, code to parse the input images, code to adjust the inputs.”- Sqrt of Pi
Rick and Ava feel that they are most artistic when managing their AI.
“It’s 80% in tweaking the parameters, but sometimes I feel like being the most creative with the input.”- Rick Vito
“Depending on the output I decide what I want and I change the constraints accordingly. My creativity matters in making those decision.”-Ava Kahn
And Frank likes the post-production compilation of video.
“Video needs a lot of tweaking before the results are good, and that is the part I like best.”-Frank Poncelet
I am not sure I have looked at enough generative art to recognise quality on sight, but after speaking with these 6 creators, I now know how good AI artists speak about their work.
