AI-generated image of an airline pilot. Generated with Midjourney version 4.

You Won’t Believe These Images Were Generated by a Computer

Reading Time: 4 minutes

Since the introduction of GANs (Generative Adversarial Networks) in 2014, AI Image generation has escalated at an alarming rate. Recently, systems like DALL·E and Stable Diffusion have put AI image generation into the hands of the public. Users can generate amazing images from the input of simple text prompts.

When Did it Start?

Artist Harold Cohen developed the earliest notable AI image-generation program.1 Cohen named the program Aaron. Accompanied by a drawing machine, Aaron created its first drawing in 1973. In the following decades, Cohen updated and tuned Aaron with new algorithms.

“In the last few years of his life, Cohen once again changed tactics. His relationship with Aaron would no longer be one similar to “creator and creature” – a creature who creates for the creator – but rather one of cooperation.”

Studio International. Delving into coding: the art of Harold Cohen

Cohen died in April of 2016. Aaron was retired shortly after. Although Cohen didn’t live to witness the full potential of AI image generation, his work was a precursor to the incredible technology we see today.

Harold Cohen holding a Turtle Robot at the San Francisco Museum of Modern Art, 1979 with a computer-generated, enlarged and hand-coloured temporary mural in the background. Courtesy Harold Cohen’s archive
Harold Cohen stands in front of a piece created by Aaron. Source: www.studiointernational.com

How Does it Work?

GANs

One of the most pivotal frameworks for AI image generation was developed in 2014, called Generative Adversarial Networks (GANs).2 GANs rely on two primary models: a generator and a discriminator.

The generator is trained to generate new examples based on an image dataset.

The discriminator looks at the images from the generator and uses machine learning algorithms to determine whether the image is “real” or “fake.”

Put simply, the generator aims to improve its output until it can fool the discriminator.

To see the uncanny ability of GANs, visit www.thispersondoesnotexist.com; refresh the page a few times, and you’ll see what appear to be real photos of actual people. In truth, they are images generated by GANs.

Diffusion Models

Diffusion models were introduced in 2015.3 The programs take an ordinary image and gradually apply gaussian noise. Afterward, the computer performs calculations to reverse the noise to approximate the look of the original image. This process can be guided so that the computer creates entirely new images from gaussian noise.

Example of forward and reverse diffusion.
Example of forward and reverse diffusion. Source: nvidia.com

Magic3D from Nvidia takes AI generation further by generating 3D models from text prompts.4 The meshes come complete with colored textures.

A poison dart frog rendered as a 3D model by Magic3D. Source: arstechnica.com

While AI image-generation technologies create extraordinary results, rapid development has left little time to address ethical concerns. In a future installment, we’ll discuss some of these concerns and what artists say about them.

Check out these images generated by Midjourney:

Including camera settings in a prompt can lead to strikingly realistic images that look like photographs.

Prompt/Command: “black and white portrait photograph of an airline pilot, taken with 105mm lens, ISO 800, f/4, 1/200th –testp –ar 2:3″ – Upscaled (Beta) by @jo5hkn

Midjourney has a ‘remix’ feature that allows users to create variations on an image with prompt adjustments.

Above, the first image is the original. The second image is a ‘remix’ of the first with the addition of the text, ‘digital artwork,’ to the prompt. For the third remix, we added the word ‘photograph.’

Learn more about Pacmin Studios custom display models or contact us for a custom model quote.

By Joshua Knopf

Joshua Knopf is a Production Expeditor at Pacmin Studios. In addition to mixing and matching colors for silkscreen printed decals, Josh writes creative content for our newsletters.

References

  1. Delving into coding: the art of Harold Cohen, Studio International. https://www.studiointernational.com/delving-into-coding-the-art-of-harold-cohen-aaron-computer-generated
  2. A Gentle Introduction to Generative Adversarial Networks (GANs). https://machinelearningmastery.com/what-are-generative-adversarial-networks-gans
  3. How diffusion models work: the math from scratch. https://theaisummer.com/diffusion-models
  4. 3D for everyone? Nvidia’s Magic3D can generate 3D models from text. https://arstechnica.com/information-technology/2022/11/nvidias-magic3d-creates-3d-models-from-written-descriptions-thanks-to-ai