Do artists using DALL-E and other AI art generation apps do more than just click a mouse?
They express themselves creatively by selecting outputs and engineering prompts.
Within only a few short years, the quantity of works created by so-called “AI artists” has exploded! Large auction houses have sold some of these pieces for eye-popping sums, sending them to prominent, carefully curated collections around the world.
How does AI Artwork Generation Work?
As picture creation technology has gotten more effective and easier to use without coding abilities, the general public has begun to embrace AI art, which was once pioneered by a small group of technically savvy artists who had already included computer programming into their creative process.
The advancements in computer vision, and the study of creating algorithms that can process meaningful visual information, have provided a windfall for the AI art movement.
The main characters in this tale are generative models, a type of computer vision algorithm. In order to learn to encode the statistically significant aspects of images, generative models, which are artificial neural networks, can be “trained” on enormous datasets including millions of photos.
After training, they can create entirely new images that weren’t in the original dataset, usually with the help of text prompts that specify the desired outcomes in great detail.
The Debate of AI Art
A lot of serious artists have been fascinated by this method, despite the fact that the images it yields have lacked the cohesion and detail of more traditional methods until quite recently.
However, a new model developed by the tech firm Open AI and codenamed DALL’E 2 has proven to be astonishingly consistent and relevant in producing visually striking results in response to nearly any textual cue.
As long as the desired impact is clearly described in the prompt, DALL’E 2 can even make graphics in certain styles and imitate famous painters very effectively. Craiyon (formerly “DALLE small”) is a free, open-source alternative to DALLE mini.
Questions like “is AI art actually art?” and “to what extent is it really made by AI?” aren’t new, but they are raised by AI art’s coming of age and are interesting nonetheless. Such concerns are reminiscent of those voiced before to the development of photography.
A camera’s “capture” function makes it possible for even someone who has never painted before to produce a photorealistic image.
Now, all it takes is the click of a button to trigger a generative model that can render images in any desired art style, from any given scene.
However, cameras and AI cannot create art. And it’s not only the people. Artificial intelligence (AI) artworks are created by human artists who, in addition to other tools, employ algorithms in their practice.
While it is encouraging that these technologies have made it easier for people to express their creative sides, it is important to remember that good art still requires a lot of hard effort and dedication.
Generative models, like any new technology, alter the creative process in fundamental ways. In particular, AI artworks provide new dimensions to the concept of curation while further blurring the boundaries between selection and originality.
There are at least three curatorial acts that can be performed while creating art with AI.
The first is the least creative, and it involves selecting which outputs to display.
Despite the fact that an infinite number of images can be generated by any generative algorithm, not all of them will be considered works of art. Photographers are highly familiar with the curating process because they sometimes take hundreds or thousands of photos from which only a small fraction are displayed.
Photographers and AI artists, in contrast to painters and sculptors, must contend with a plethora of (digital) objects, the curation of which is integral to their creative process.
The practice of “cherry-picking” the best results from a dataset is often frowned upon in the field of artificial intelligence research since it can be used to artificially inflate a model’s apparent effectiveness.
However, cherry-picking can be an effective strategy when creating works of art with AI. The sheer act of elevating certain results to the rank of artworks may be a manifestation of the artist’s aims and creative sensibility. Fundamentally having the eye for a good and appealing composition.
For a second, photographs may be curated even before they are created. However, in the context of artificial intelligence (AI) research, “curation” is commonly used to refer to the process of creating a dataset used to train a neural network, rather than the selection and presentation of artistic works.
Important as it may be, this task is necessary since a network may not learn to represent required properties and perform adequately if the dataset is badly built.
In addition, the network is predisposed to duplicate or even amplify any bias present in the underlying dataset, which could include, for instance, negative prejudices. “Garbage in, garbage out,” as the adage goes.
As with every kind of art, the adage applies to AI artwork, with the exception that the term “trash” now has an aesthetic (and subjective) connotation.
In order to create his groundbreaking AI artwork Memories of Passersby I (2018), German artist and early pioneer of AI art Mario Kinglemann painstakingly collected a dataset of thousands of portraits spanning the 17th through the 19th century.
Next, he put this data to use by training generative algorithms to generate an endless supply of new portraits with the same general style, which he then presented on two screens simultaneously (one for female portraits, one for male portraits).
An example of AI artwork that does not require human curation of the final product. However, it couldn’t have been conceived of without the careful curation of the training data. “Bias” is a virtue in this context: The artist’s aesthetic preferences and taste substantially influenced the dataset, and those preferences and tastes are mirrored in the final artwork, albeit distorted by the algorithmic generative process.
Another innovation made possible by recent advances in generative algorithms is the generation of images from a plain language description of the desired outcome. This method of instructing an algorithm with text prompts as opposed to sampling its outputs at random is known as “prompting.”
Take a look at the picture that goes along with this article: Multiple pictures in the collage were generated by the generative AI model DALLE 2 in response to the prompts “an AI image generation algorithm, conceptual art,” “collage using images formed by a generative AI model,” and “an artist curating artworks produced with an AI algorithm, conceptual art.”
Words as input to a generative algorithm can simplify and direct the creative process. If one’s vision can be described explicitly, then perhaps less time will need to be spent curating results.
Prompting, however, is not a panacea that reduces the significance of creative work. Actually, it’s closer to a brand-new creative ability. Some experts in the field of artificial intelligence go so far as to use the term “prompt engineering” to explain the process of developing effective prompts to achieve specific goals.
When it comes to innovative applications of AI, prompt engineering is more art than science. It’s been likened to incantation or alchemy. It’s not enough to have a clear idea of where you want to take the end products; you also need a sense for the magic words that will let any given algorithm understand the nuances of particular styles and topics.
Herein lies the third and, perhaps, most original type of curation presented by AI art: the deliberate creation and collection of individual prompts (or prompt fragments) that elicit the intended outcomes from an algorithm.
Prompt curation provides an alternate route to establishing a unique aesthetic as the usage of pre-trained algorithms such as DALL’E 2 begins to negate the necessity for dataset curation.
Like typical museum curation, it pairs visuals with words, but the tone is less scholarly and more poetic. Writing prompts are like a form of art criticism, in that they can range from the concrete ideas of (“A woman riding a bicycle down a wintery country lane”) to the ethereal idea (“the eternal existence of the light of god”).
In either case, prompts add a fresh dimension to the interpretation of works of art. While some creators enjoy discussing their inspirations with others and even incorporating them into the titles of their finished works, others choose to keep their ideas private.
In the creative process, a feedback loop is formed between the selection of inputs and the selection of outcomes. It is possible to experiment with a set of prompts, learn what kinds of photos they can generate, and then utilise that information to iteratively improve the prompt while simultaneously selecting the most interesting results.
One can continue this loop forever. It’s evocative of how classical artists would experiment with different takes on a theme, like Picasso’s The Bull (1945) lithograph series, which represented a bull in varying degrees of abstraction.
An important distinction is that the randomness of the generating process ensures that no two prompts will ever yield the identical outputs, and that even little changes to the prompt can have a surprisingly huge effect on the results.
There has long been a porous line between the roles of artists and curators. It wasn’t until the 1960s that curatorship was acknowledged as more than just a custodial job, one tasked with maintaining and presenting a museum’s collection of artworks for public consumption.
In order to present a group of artworks in a novel light, a curator may embrace a specific theme or point of view. Star curators like Carolyn Christov-Bakargiev and Hans Ulrich Obrist take an artistic approach to their profession and have shaped the way we talk about art and curatorship today.
On the other hand, artists like Marcel Duchamp played a vital part in updating the exhibition medium by curating landmark events themselves. The act of curating can be seen as an artistic endeavour in its own right, allowing for a highly individual statement of aesthetic preference.
The development of generative algorithms introduces new curatorial gestures that channel the artist’s aesthetic sensibility at various phases of the creative process, thereby expanding the scope for collaboration between the two disciplines.
These curatorial considerations of AI art may someday permeate museum or online exhibition curation approaches. Institutions showcasing AI artwork, for instance, may have to make a call on how much detail to reveal about the datasets used to train the algorithms that created certain works.
The fact that portraits from the 17th through the 19th centuries were included in the training dataset is mentioned in the catalogue note for Memories of Passerby I at Sotheby’s, which is helpful for situating the work within its artistic canon.
Some curators may choose to include and comment on the artist’s communication of the prompt that inspired their work if they feel it adds context and significance to the exhibition.
In keeping with the concept of the curator as (AI) artist, one could also imagine an exhibition in which classical pieces of art are chosen for display based on the degree to which their descriptions share commonalities (see Google Arts & Culture for similar experiments in digital curation).
One thing is certain: the rich ground for new kinds of creativity provided by technological advancements stemming from AI research will continue to influence artistic creation and curation in interesting and unforeseen ways.
