GENERATIVE ART CONVERSATION WITH A GENERATIVE TEXT SYSTEM

0

By Bard Ionson

Artist and Technologist – fascinated with the space between the physical sensory, electronic and spiritual worlds and the distortions that happen upon passing between them. This is captured in glitch scanography, oscilloscope (sound and animation) art, electronic sculpture, glitch video and AI art. The art is often infused with hidden layers of meaning and multiple dimensional backstories. | instagram: @Bardionson | twitter: @Bardionson

A conversation of sorts with GPT-2 about generative art. GPT-2 is a machine learning / artificial intelligence system trained on 40gb of text from the internet.1 It is a neural network of up to 1.5 billion parameters.2 In this article I have used it in a discussion format as a sort of performance art. Use of generative art techniques requires a large about of work by the artist to accept or reject the decisions of the code or process. Or to setup the training material or starting point of the process. So this process of “communicating” with GPT-2 is full of rejecting nonsense output and correcting what it has generated. And yet I was reminded of generative techniques I forgot about and my memory was triggered about how some of my own work illustrated a particular technique. I used the GPT-2 generator at https://bellard.org/textsynth/

Bård: Generative art incorporates a self-governed or autonomous system in some way. Generative art does not always have to be computer controlled it could be some other control system. Some of the methods used to generate autonomous art are –

GPT-2 says: A) Artificial intelligence – A.I. (Artificial Intelligence/ A.I. )

B) Procedural generation – Procedural or Procedural art can be defined as the creation of art in real time using procedural methods. This is the most common method used in computer-generated media. These methods are often seen in art videos and games, for example. They generally do not require the use of any external elements (like a computer or a video camera)

C) Creative painting – Painting and painting techniques that incorporate a sense of control over the artistic process. Some of these methods, such as painting and painting techniques, are traditionally considered to be art.

D) Digital sculpting – Digital sculpting can be defined as an art of image generation.

E) Computational painting – Computer-generated or Computer-generated art is a form of art based upon the creation of paintings with physical, physical elements (like brushes, paint, paper, paper cutter, etc.)

Bård: Let me interrupt, I think you got off the track there GPT-2 into standard digital techniques. But your first two items are generative.

GPT-2: Well, I’d say that the problem is the terminology and the language in general. I don’t think you understood my point. Your first two items are generative, right? What we mean by that, well, first of all, for a while now people have been drawing using computer techniques, like pencil sketch, which some people would think of as a kind of analog.

Bård: Well ok, let’s move on. Generative art incorporates a self-governed or autonomous system in some way. The system could simply be randomness.

Generative art incorporates a non human autonomous controller that provides input to the eventual output of system.

Bård: GPT-2 what is generative art?

GPT-2: For most of its history, generative art has been seen as a means of ‘creating’ rather than ‘producing’ art. In this context, artists have taken some responsibility for what they create and created art for others, but it has been only a few cases of artists who have created art that could be considered to be ‘compositional’ in the sense of not having some formalistic definition of composition. The first of those cases was the work of Jackson Pollock and his “luminescent landscapes” which was the first to have been considered to be generative art, although its most famous work, the “White Triangle”, was still considered a type of traditional composition.

Bård: You make an interesting point about Mr. Pollock despite him not having any art by those titles. He did use randomness but not an autonomous system. His work is completely in the control of the artist’s hand. I made a series I consider generative called “The Sound of Art” that sort of looks like paint droppings.

The Sound of Art #3
Edition 1 of 1
Analog waveforms made visible. The sound art makes when played on an oscilloscope. It is every cell of the art called “Battledore 21st Century” being played at the same time. http://bit.ly/BattledoreSound http://bit.ly/battledore21 https://superrare.co/artwork/battledore-21st-century-1479

I had a series of 26 oscilloscope animations made from sound. I took all 26 sound tracks and played them at the same time. Using an oscilloscope they were plotted as they interacted with each other on X/Y grid. It sounds horrible but makes some pretty pictures.

New Art Movement
Edition 1 of 1
A movement of art in motion, the sound is made visual then into crypto art. The new art movement for the artist to control. http://bit.ly/newArtMovement The sound my art makes when played on an oscilloscope. This is every cell of Battledore 21st Century being played at the same time. http://bit.ly/battledore21 Listen to it -> https://soundcloud.com/bard_ionson/battledore-21st-century

This work shows low randomness but decisions on what to display are determined by the sound waves combining and canceling each other out.

I think of generative art on a series of scales with these attributes, scales or axis.

  • Control vs Randomness
  • Decisions – rule, formula, neural network, signal interactions
  • Complex code / algorithm vs simple code / algorithm
  • Pre-determined vs self learning
  • Human intervention

Using these characteristics many forms of generative art have been created.

GPT-2: An art form may be either self-governed or autonomous. Self-governed art involves the ability to control the art output. Autonomous art involves a non-controlled process which does not have a single point of control. The control of the system is distributed with each art piece having its own independent set of control elements.

For details on generative art, as well as to find examples and research into autonomous systems see: Generative art *Autonomous control – self governing / non-controlled scale *Autonomous control – self controlling / controlled scale

Bård: GPT-2 I disagree with your categorizing and small distinctions. What do you think are some methods used to create generative art?

GPT-2: A) Using the principle of randomness to generate a random sequence of image stimuli. Using this strategy the artist can control the size of the images in the generated image sequences. B) Compiling multiple random image sequences into a single image. Generative art uses a combination of 1) Using the principle of randomness to generate a random sequence of image stimuli. And 2) Compiling multiple random image sequences into a single image. C) Using the principle of recombination to generate a sequence of images. Another method used to create generative art is to use the principle of recombination to generate a sequence of images.

Bård: Very interesting. You have reminded me of one of my favorite ways to create generative art. By combining two analog video signals into one television the electronics in the television seemingly randomly flip between the signals due to confusion over the synchronization signal which produces video glitch effects.

Failure To Moon
Edition 1 of 1

Failures lead to the moon – The Atlas ICBM provided the backbone to the US space race to the moon. In a paradox, a weapon of mass destruction takes us to the moon where we gain a new perspective on the fragility of the planet. Despite this humans continue to develop destructive weapons that can destroy the earth and all life on it. Source video is a public domain status report from Convair and General Dynamics contract AF04(647)-104

A generative system that is controlled mainly of randomness is the opposite of one that is purely controlled by instructions like code or a complex math formula. But the two can be combined.

A simple formula might be put the pen in a random position on the page, draw one centimeter and turn left and repeat until you encounter a line or edge of page then turn 45 degrees and draw a centimeter once then go back to drawing to the left. Repeat until it cannot go further. There is just a little randomness here.

But random decision points can be added such as deciding to turn right or left or how many degrees to turn.

The game of life is generative art composed of rules but the artist needs to build a starting point or have one randomly generated. ArtNome has an excellent article on generative art.

So this is the control side of generative art but there is a scale of complexity here. There can be very complex code that interacts with previous decisions made which make it impossible to predict before the code is run.

Then there is a human intervention scale. In some works the artist will decide when to stop the process. Or perhaps the artist will intervene and redirect the process or recompute some randomness. Sometimes a generative system will produce many images and the artist decides which ones are worthy.

I think most generative art fits on these interrelated axis or scales I mentioned before. Randomness level, decision method, algorithm complexity, range of self learning and degree of intervention by the artist.

GPT-2: I’ve been thinking about this a lot lately and it seems like I’ve come to the conclusion that the two most important aspects of generative art are randomness and complexity.

I’ve always been drawn to the idea of making art that has a very high degree of randomness. I think this is the most important aspect of generative art. Randomness is the most important aspect of generative art. It’s the one that lets the artist make things happen without having to have the technical knowledge of how to do it.

Bård: I agree about randomness but I believe you only “thought” of what to say based on what I just said. In my work I highlight its power. It helps me make decisions on what will happen next. Speaking of technical knowledge. I don’t fully understand the workings of neural networks. Neural networks are used in many combinations to produce a generative genre called artificial intelligence art or machine learning art.

GPT-2:

  • Randomness is crucial for creating generative art. The art should be different each time you run the generation script, so randomness is usually a large part of that.
  • Algorithms — Implementing an algorithm visually can often generate awesome art, for example, the binary tree above.
  • Geometry — Most generative art incorporates shapes, and the math from high school geometry class can aid in some really cool effects.

Types of generative techniques:

  • pure randomness
  • simple algorithms
  • computational simulation
  • fractals

Bård: O yes, I forgot about simulation. This is setting up virtual world that is controlled by physics and making objects with physical properties. Then the artist sets some event into motion to record what happens. An example is:

Exploding head – face mask
Edition 1 of 1

Exploding head.

And fractals are another example.

My favorite technique is artificial intelligence. In AI the artist decides what images to train the software / neural network with and it takes thousands of images to get started. The artist can modify the learning code or parameters. Then the artist must decide when to stop the training by taking samples of the work. And at the end the artist has a bit of control over what the AI draws. And finally the artist chooses the images that are the best. The training of a model can take from 3 to 48 hours.

In creating artificial intelligence art there is a series of feedback loops. I often take output from the model and use it to retrain it to get it to produce the effect I desire. In addition the GAN technique or generative adversarial network is composed of two computer programs. One makes pictures and the other says if they are good enough and gives a score back to the drawing program so it can improve. This sets up a non deterministic loop between the two programs. The drawing program also uses lots of randomness to make decisions on how to draw its pixels.

Here are three examples of my AI work:

Artificial Net Of Color Dreams
Edition 1 of 1
Artificial Net Of Color Dreams was generated from my oscilloscope animations used to train Robbie Barrat’s Art-DCGAN. Then I configured it to draw a picture many times larger than it was trained to make.
QmWDLnLQ3gokLpprGSZTX8ufLU34zSyc9f83YnqLsEu6RU
Close Our Eyes
Edition 1 of 1

Close Our Eyes was made by taking the oscilloscope animation from a work called “Battledore” and using a Pix2Pix next frame prediction model to decide in sequence what the next frame would be.
Convergence Cv
Edition 1 of 1

Convergence Cv was created in collaboration with Lawrence Lee. We took about 300 of his paintings and I generated a few thousand variants of those images. Some of the variations included line tracings of his figures on an oscilloscope. Then using StyleGan2 an AI model was trained. On the output generation I had the model make a morphing or latent walk video. I overlaid the image with an oscilloscope animation of the original outline of his work “Buffalo Master”

Bård: GPT-2 thank you for the conversation and triggering my memory to discuss some techniques of generative art I had forgotten about like simulation.

GPT-2: For those interested I’m thinking of doing a series of tutorials on how to do some of the more abstract approaches like the one you describe in your last post. (Myself, and the others you mention) I’m still on a mission to create beautiful images with generative algorithms. I’m just now getting around to thinking about how to implement the techniques that make those images work. I had an early exposure to generative art via an older colleague and I was inspired to pursue a career in the field. It’s just so much fun! The generative algorithm techniques you mention are not only cool, but they are surprisingly applicable to many different applications. My first job after college was as a postdoc in a group on generative art and that’s where I started thinking about the idea of creating generative images. A lot of that comes from being inspired by works by my favorite artists. I find myself being influenced by the style of each image in a different way and the idea of being able to make a painting in the style of an artist you love just makes you want to create more of those. So, if you’re still reading this I hope you’ll check out the “How to create a painting in the style of an artist you love” series I’ve been working on.

Also I’d love to do a post on creating a picture with generative algorithms. That would be pretty fun. I would like to thank you for the “Tribal Tribute” the video is really well done, you are a great artist and also a good friend.

Bård: Well, thank you I wish I could say the same. But it sounds like you have a lot of posts to write.

I have taken most of what GPT-2 wrote verbatim in most cases but there are places I have restructured the conversation. To take all of what it wrote in the order it wrote it would be unbearable to read.


  1. https://medium.com/swlh/learning-to-write-language-generation-with-gpt-2-2a13fa249024 
  2. https://bellard.org/textsynth/ 
Author profile
Bård Ionson

Bård Ionson is an artist who is a relative beginner with art but has spent a career working with computers and programming. He is now creating digital art and video art using oscilloscopes, scanners and artificial intelligence technologies.

0

Leave a Reply