Artificial Intelligence Today Can Paint (Almost) Like an Artist. Some People Think That’s Stealing.
FlamingCraze sits curiously, staring at the laptop screen in his dingy room. Midnight has just set in and it’s getting quiet and rustic outside. The result of the past 6 hours of anticipation is in front of his eyes. His Stable Diffusion model has been trained.
Like many other people who follow this hobby of generating AI art, he has fewer qualms about the nature of his work. He wanted to be a digital artist for a long time. During the pandemic when he was locked in his home like everybody, he enrolled in online courses for the same and started practicing using his digital tablet, now dusty and sitting at a corner of his desk. It was a long-drawn process. However, he doesn’t seem to have made much progress.
“Learning the tools and techniques was hard. I didn’t have any formal training in painting. I spent weeks going over the basics and produced a few works. They were simple toy projects that beginners do. However, it was not something that I wanted. I have seen so many artists on Instagram, their work is so cool. It will be years before even I can come up to their level” he exclaimed.
He knows learning a new skill requires patience and time. In this age of getting everything at your fingertips at the touch of a button, it’s not surprising to see people lose their passion as soon as they see diminishing returns from their learning.
So, when he came to know about AI-generated art, it immediately caught his attention. “You can just describe what you want, and the model will just paint it. This is amazing.”. You can see the fascination in his eyes as he showed me an example — “Rain on Mars”, he typed that on a text box, and within a moment, you can see a picture, a red background, some black hills in silhouette here and there, and some white particle-like things which seem to fall on the ground.
Since then, he has been part of this subculture of folks who create AI-generated art and post it online. He shows me his Instagram page. “I have gained around 3000 followers in the last 3 months. Many people like the art and have complimented me. I am looking for ways to improve it and get better at it”.
The irony is, he is not the one who is producing the art. Will it be his effort if it gets better?
Like FlamingCraze, I got in touch with some other so-called AI artists in order to better understand their work. Most of them go by pseudonyms on forums such as Reddit and Discord. These, however, seem to be a lively and welcoming community for beginners and experienced alike. Their posts are filled with tips as to how to get better results from these AI art tools, and unlike most places of creative expression which seem to become a dumpster fire of negativity sooner or later, these ones contrastingly had a good positive vibe.
The Complexity of Generating AI Art
“You cannot just type any text and hope to get results,” says one of these AI artists who goes by the name of RiverGoat. “The AI model understands natural language text, but still, we will have to tweak the phrases here and there, in order to get the result we want. It’s mostly a trial-and-error process, you must keep tweaking the words until you find something you are looking for.”
When you browse their forum, you will come across a large bouquet of creativity. It’s akin to visiting an art gallery, but the styles and concepts are varied. Many of these folks seem especially interested in copying some anime art styles, some like to bring to life a fantasy world. Like a child in a toy shop, I had a look of wonder at all these pictures, which surprisingly, have never been touched by humans.
I couldn’t help but ask myself, will the prophecy of AI replacing humans in their jobs, the jobs of artists in this case? Is it coming true?
“It’s not just as easy as that. Sometimes we will have to rent GPUs by the hour, and train models for a long time on them. The results posted on Reddit are the ones that we find decent after hours of our effort. Most of the images produced by these AI are not perfect. The limbs of the characters get disfigured, the face gets discolored, and so on. Sometimes we take the model and have to train more images on top of that, just to get some good enough results” says RiverGoat as he explains to me the limitations of these systems further.
An Unexpected Turn of Events
Not everyone is happy with AI generative art though. Especially the artists, who earn their bread and butter from this trade.
Hollie Mengert woke up one day, out of the blue, by getting an email that someone had copied her style and created an AI model. She is one of the most well-known and successful digital artists out there, who had created works for companies such as Disney and Penguin Random House. Like most creators, she was not pleased.
Someone took thirty-two of her pictures and trained an AI model with them. The result — the model had effectively learned the style of her paintings, and now can be used to generate pictures, as if she herself has drawn them.
Apart from the ethical question of whether that is moral, Hollie Mengert has raised the issue of the legality issue of training the model with copyrighted works. However, like many artists, she didn’t hold the copyrights of these works either, it was owned by the companies that had commissioned her to draw them.
The AI artists had a different take on this issue altogether. “You cannot copyright a style”, one of the comments pointed this out on their forums. Some argued that this falls under the fair use policy.
“When one trains an AI model, they are not storing each pixel of the image in the model. The model just tries to find patterns from the images and saves them. This is no different than some artists getting inspired by other artists and trying to inculcate some of their styles as their own. The training images may be 100s of GBs, however, the models themselves only weigh a few GBs.” — FlamingCraze explains this, as he shows a Google Colab notebook, open on his laptop screen, a platform where one can train these AI models.
The concerns of the actual artists were however more than just technicalities of the technology. Someone had used Hollie Mengert’s name in the model, and she hadn’t approved of that. The AI artist who had trained this model, however, complied with her request and removed the name from the model. He further added in the description — not in any way affiliated with Hollie Mengert.
Some do not share the same empathy though. “Adobe is adding image generation to their creativity suite. So will some other companies like Microsoft. I think it’s a sign of the times that artists should adapt to this change.”
More Than Just Art
“I think what we are doing is imaginative. People have created models with the art styles of long-dead artists such as Anders Zorn and Van Gogh. How will these artists paint some modern-day scenes? Now we can know. In a way, we are keeping their legacy alive”.
It goes without saying that for these AI artists, AI-generated art is more than just training the models and generating art. You can feel a sense of community and belonging, with each one helping out the other. They create these models and share this work openly with each other. The HuggingFace platform where many of these models are hosted boasts a collection of nearly 800 “concepts” to date.
One can see this collection will only be growing. And so will be the questions like — If an AI learns to paint by seeing the works of some human artists, is it stealing?
