The Audiences for AI-generated art: imagining the future of creativity

The Audiences for AI-generated art: imagining the future of creativity

  • Tech

30th November 2022

Even ten years ago, the idea that anyone could wake up from a vivid dream and immediately be able to create images of it by simply telling the computer what you saw would be preposterous. Recent leaps in technology have eliminated the need for long-winded stories about dreams and replaced them with dazzling pictures of, well, anything the imagination can muster.

We find ourselves on the edge of an AI art revolution – and we’re using Pulsar to capture the beginning of it.

From art to commerciality to NFTs, advances in AI-generated imaging and public access to it have set the social conversation snowballing.

It’s clear that people are talking more about AI-generated art and the prompts that create it. But what exactly are we talking about? And how does it work? 

Different AI models work in their own ways. Put it in simplest terms, AI image generation models are fed images from the web along with their text descriptions and then use that data to make an image based on what the user tells it to.

The text description a user gives a model to create a brand spanking new image is called a ‘prompt’. Prompting can be a hard task – models respond differently to descriptions based on what they’ve learned. For example, rather than including the phrase ‘high quality’ users get better results putting ‘award-winning photo’ or ‘Canon EOS R6’ after a prompt to increase the chances of a high-res image output – images online with those phrases in the description tend to be of higher quality.

When we look at the broad topics of conversation surrounding AI art online, we can see that the discussion around prompting roars louder than the conversation around ethics – discussing the validity of the AI-created images as art in their own right. This conversation dominates news conversations around AI art, but global social media audiences prioritize creativity and prompt discussion.

It’s clear that different audiences are using AI imaging for various reasons – and all AI imaging models have their strengths and weaknesses. We examined the conversation leaders in AI imaging discussion online – let’s take a look at what each sub-audience is posting about.

It’s clear to see that AI art audiences online skew towards creativity and techiness, with some audiences like 3D Art Devs and NFT Collectors falling across both of those brackets.

When we look at our conversation-leaders audience breakdown, we see that prolific AI art posters on Twitter give even less importance to discussion ethics than the total online AI imaging conversation. These experts are one step ahead of the conversation, bypassing questions of ethics and simply ‘doing’, through creating and prompting – as well as being part of the burgeoning AI art market.

So who are the main players on the AI art generation stage?

The most impressive models (such as DALL·E 2 and Stable Diffusion) utilize ‘diffusion’, a technique where the AI learns each training image and accompanying text by destroying it through progressively covering it with random pixels, then putting it back together again in reverse to rediscover its meaning. It takes each step through billions of variables, performing further wizardry that…we’re not going to get into here.

OK, that’s as technical as we’re going to get – but for those interested in the (terrifying) math, have a look here.

But technical questions are only the tip of the iceberg. Search data reveals that online audiences are asking big, existential questions about AI art.

Within this discussion, there are several perspectives. A prominent perspective comes from those that argue that by developing the language that is used to generate an image, a person is indeed creating a valid piece of their artistry, and is just using a new medium to create their art.

Creativity and imagination are at the heart of AI art – any person can create their dream image without lifting a pencil, opening any Adobe software, or buying a top-end camera. The ability to create whatever the imagination desires is at the fingertips of a formidable group of people: geeks. So what are audiences making images of?

Here we compare the top styles and characters/subjects of AI-created art online. Both subject and style are almost equally important for users, with Anime proving to be the most popular aesthetic, and Disney being the most favored for choice of subject.

It’s important to note that AI models have limitations on what can and can’t be created using their data. For example, most models ban the creation of any kind of NSFW image. (This doesn’t stop users from trying to get around that, however. When we first launched our Pulsar TRAC search into AI imaging, a staggering over 10% of results were users creating and discussing how to create images of large-breasted women around the NSFW ban.)

It’s not surprising to see so many animation-based types of characters and art being created by users – after all, AI imaging is infamous for its failure to fully capture an accurate human body.

Let’s look at the difference between the types of artwork created across the top AI art creation models.

The prevalence of different styles across different models indicates not only that those models excel in those styles, but that they attract different types of audiences.

We see this again when we examine AI-generated images’ subjects. Prompters are using different platforms to create AI-generated art of their favorite characters.

Of course, being able to replicate any style, subject or character brings up questions of copyright. Artists’ arguments around intellectual property, authorship and validity frame AI art as nothing beyond ‘theft’, removing profits from traditional and digital creators.

The commerciality of AI-generated art is still being figured out. It will undoubtedly have wide-spanning commercial applications. The header image for this blog was made using Nightcafe. But for now, conversation is mainly limited to its validity as ‘art’.

One of the arguments against seeing AI art as art is that the effort needed to create AI imagery is highly limited in comparison to the endeavor needed to create ‘human’ art, and that AI pieces can perhaps even be exploitative of professional artists trying to make a living.

Currently, an artist’s style cannot legally be subject to copyright (though certain types of derivation could possibly be seen as theft).

Some argue that AI imagery is indeed art, and that their position is ‘endorsed’ by one of the most famous art galleries in the world, with the MoMA in New York staging an AI-generated exhibition.

Unsurprisingly, we see social media outcry from those that vehemently disagree.

Weird Dall-E Generations (@weirddalle) is a Twitter account with 1.1 million followers that shares amusing and weird AI art. However, this account, one of the most prominent ai-generated image posting accounts, has nonetheless ‘come out’ as not believing ai-generated art is ‘real ‘art’. 

At every stage of technological innovation, artists innovate their production methods – and along with it comes discussion as to what it means for ‘real’ art. Digital art has existed ever since technologies made it possible to create and paint digitally and is still subject to scrutiny in the art world and popular culture. 

The 15th century saw religious luddites fear the invention of the printing press, worried of it destroying the sanctity of the scribed word. Similarly, technophobes in the animation industry in the 1990s voiced woes about the invention of CGI. These days computer-generated animation is now praised as a beautiful form of Oscar-worthy art – indeed we found Pixar to be among the most popular styles of AI-generated art.

We’ll let you decide the next bit. Below are two sets of images – one set created by hobbyist artists at Pulsar using AI in the style of Van Gogh and the other painted by a hobbyist artist in the style of Van Gogh. We’ll leave you to it…


 
To stay up to date with our latest insights and releases, sign up to our newsletter below:



This article was created using data from TRAC