The world of AI-generated art is gaining more and more traction. For example, artificial intelligence has proven that, given a text prompt, it can create images based on the style of a particular artist. This is an example of generative AI, which is changing the creative landscape.
However, it is not without problems. A Business Insider article described how Greg Rutkowski, previously known as a specialist in creating images of Dungeons and Dragons scenes, is struggling with this. Typing the prompt “Dragon battle with a man at night in the style of Greg Rutkowski” into the website Stable Diffusion, today produces tens of thousands of images in his same style.
Stable Diffusion is not the only one with this type of technology, OpenAI’s Dall-E and others provide this generative AI as well. With AI generated art ramping up in popularity, new regulatory guidance has been unable to keep up. For example, there is still confusion around whether or not AI-generated art infringes on copyrights for example. This article will discuss AI-generated art in detail. We’ll also cover what AI-generated art is and the disruptions that generative AI is causing.
What is AI Art? (1:13)
So, what is AI-generated art? I think to answer this question we need to take a step back and understand more generally the concept of Generative AI. The concept is similar no matter what media you’re generating.
Generative AI uses images, videos, audio, text, and 3D models as training data and can then provide back new, creative forms of the same or different media. One of my favorite examples of generative AI is a tutorial on “How to prepare for a spaceship adventure as a cat.” The video including the human avatar was fully generated with a text prompt. I tried the same prompting and received a similarly absurd video shown below.
In the case of image generation, the two most common types of inputs are other images or a text prompt. The first is known as image-to-image generation. Another even simpler technique is known as text-to-image generation. In this case, you don’t even need to source your own images, but rather just come up with a simple text prompt.
What are some of the most popular generative AI image tools? (2:09)
The first category are the image-to-image AI art generators which take images as input (as opposed to text). I’ll mention too, that both of these platforms can do text-to-image generation as well, though the image-to-image generation facet really makes them unique.
Image-to-image AI art generators
NightCafe (2:32)
The first in this image-to-image category comes with a cool backstory as well. Started by Angus Russell in 2019, the company NightCafe Studio came about when Russell needed to decorate the walls of his bedroom. He found that the art available to him didn’t seem personal and tailored just to him. Thus NightCafe was born. It originally began as a neural style transfer app in 2019. Neural Style transfer takes a content image as well as a style image as input. The technique then blends the content image together such that it still looks like the content image, though in the style of the style image.
The platform still allows you to use the neural style transfer generation tool as well, albeit with their older interface. To use it you create an account, and upload both the content image and the style image of interest. NightCafe now appears to be primarily focused on text-to-image generation techniques and provides a newer interface for those tasks.
GetIMG (3:27)
Probably the coolest image-to-image generation platform is GetIMG. In addition to being able to do text-to-image prompting, with this platform you can literally create a tailored image generation model just for you. This is great for creating avatars and transforming yourself into all kinds of cool generated scenarios. It’s also great for taking a product that you might be marketing and adding it anywhere you can come up with that fits the branding.
It works by going to the DreamBooth Studio at GetIMG and selecting the model type that you’re interested in creating: choose between Avatar to put yourself into any scenario, a product shot that can take the product you’re marketing and add it to a mountain for example.
For the Model style, you have a bunch of choices, but I tend to go for Stable Diffusion 2.1 to enable the resulting model to be as versatile as possible. Next, upload images of the type that you want to fine-tune the model for. You can adjust the hyper parameters in the advanced settings as well. After clicking “train”, it will take some time so be patient. Behind the scenes it is training a new image model for you to use.
Text-to-image AI art generators (4:19)
The next category of AI art generators are the text-to-image AI art generators.
MidJourney (4:24)
The first in this category is Midjourney. When I was researching for this episode, I was blown away at the high caliber team working on this product. On their board they have the CEO of GitHub as well as Jim Keller, silicon guru who’s impacted many top tech companies. To use the software is a bit non-traditional however. To get access to MidJourney you’ll need to join a Discord server. When you have access to their Discord server, you can then join one of the newbie chats to start off. You provide prompts starting with the /imagine function to guide the software into generating images for you. As of May 15, 2023, there are over 15 million users who have joined that Discord server.
Dall-E 2 (5:03)
Next in the text-to-image category is likely the most publicized platform since it came from OpenAI. OpenAI released DALL-E 1 in January 2021. Afterwards, they released DALL-E 2 which provided 4 times more resolution. OpenAI has been deliberate in terms of how they limit the model to stay within strict bounds that they define. For training images, they only provide images that they deem suitable for public consumption. Additionally, they’ve ensured that they have removed real faces from training image.
To use DALL-E is pretty straightforward, you can access it on OpenAI’s website and you can get free credits monthly which is great.
Shutterstock (5:40)
The last generation platform that I’ll mention is Shutterstock. Now, Shutterstock might be most commonly known for creating invitations or photo albums, but they are also making the foray into AI and image generation.
To use it, you go to the generate section on Shutterstock and type in a prompt. You can then use the zoom tool if you want to essentially add more or subtract from the image where you’ll get new generated images. You can then pay for a license and download the image if you like it. Under the hood, Shutterstock’s new image generator is built on top of Dalle-2 with some additional features on top.
Who owns AI-generated art? (6:17)
So, if you generate art with AI on one of these platforms, who owns it?
One entertaining anecdote on this topic is the story of a monkey nicknamed Naruto. The British photographer, David Slater was taking pictures of wildlife in Indonesia. After putting the camera on a tripod to take a picture, one of the monkeys, Naruto, took some pictures of his own.
Technically, Naruto was the one to snap some of the photos. This raised the question, then, was Naruto the legal owner of the images? The US copyright office took their stand. They mention that they will only process claims of copyright from humans.
This is widely taken to mean that monkeys such as Naruto as well as AI (being nonhuman) cannot be submitted to be the owner of a copyright claim. So if non-humans can’t possess a copyright, then “AI-generated art has no owner” according to the Center of Art Law.
However, there are copyright owners for the images that were used to train the model initially. Lots of image generation platforms were trained on real images that were produced by real people. These artists do hold the copyrights on those images. An example of this is already playing out in the courts. The stock image company Getty Images claims that Stable diffusion infringed on their copyrights.The ongoing lawsuit alleges that Stable Diffusion improperly utilized their images as training data in their image generation software.
If the legal disaster didn’t already seem big enough, it gets worse. If it is determined that Stable Diffusion did violate copyright by irresponsibly training their model, those that utilized the platform to create images likely would be in trouble too. If that’s what happens it’ll be near impossible to backtrack and remove the images since so many are already out in the wild.
I do have an interesting thought experiment for you though. So all legal precedents aside, who do you think should be the owner of AI-generated art? Is it the humans that trained the model, the AI itself, or no one? If we take OpenAI’s DALL-E as an example, it was trained on millions of images from real artists. It then draws from this knowledge of images and uses them as inspiration to create a new, unique art piece. It actually sounds quite a bit like what humans do when they train a new skill. As humans, we’re able to go to museums or online and view art pieces. Those images rattle around in our brains, either consciously or subconsciously, influencing at least a little bit how the final product comes out. Neither an AI model nor a human artist exactly copy the art they’ve been trained on, so does it make sense that humans can own their art and AI can’t?