Creating masterpieces from AI
One of the most interesting aspects of artificial intelligence in 2023 is its ability to be creative. We once thought that AI would replace jobs such as manufacturing or self- driving garbage trucks. However, AI now appears to have significant capabilities in creative industries such as graphic design, copy editing, concept art and music. Is this the end of artists? Tavik Morgenstern, Academic Integrity Investigations Officer, explores this in more depth.
I’m reminded of a line from the movie I- Robot, where the robot is interrogated by Will Smith.1 The robot appears to think it is sentient and has feelings. Will Smith challenges the robot: “Can a robot write a symphony? Can a robot turn a canvas into a beautiful masterpiece?“ The robot, puzzled, simply looks at Will Smith and says, “Can you?”.
When I first came across AI art a few years ago it was messy. Laughable even. Things have rapidly changed in the last few months and the definition of what it means to be creative has been brought into question. Some of the imagery created out of AI is breathtaking. See an example with Image 1.
So, what happened? How did AI art become so good? Basically, a new method was developed called Stable Diffusion by a company called Stability.ai.2 Stable Diffusion works by taking in a text description as input and using its neural network architecture to map this description onto image features. It then uses a diffusion process to iteratively update and smooth out these features until it generates an image that closely matches the input text.
Image 1 created by Tavik Morgenstern using Lexica.art
How to create an image with AI
There are numerous websites you can access now that will allow you to generate AI art with ease and a lot of them are free. The paid versions will generally offer a more polished image output than the free ones, along with a license to commercially sell the image if you wish.
If you go to this hugging face page, you can try to create a piece of AI art for free here using Stable Diffusion’s original model. Hugging face is a large repository of AI software being developed around the world that anyone can access to play around with.
In the prompt area, type in a short description of the image you want. Generally, the more descriptive you are, the better the image quality. If you have a good knowledge about art or know all about cameras, you can use this knowledge to your advantage. You can also type something into the negative prompt box too of something you don’t want in the picture. E.g. low quality, pink.
Type this simple prompt: Landscape image of a forest.
You should see four versions of an image of a forest appear. Not groundbreaking. But cool in that these are all original images. Try expanding the prompt:
Landscape image of a forest by Edward Hopper.
Better and more interesting. Adding an artist name helps Stable Diffusion understand the style you want in your prompt image. You can get interesting results if you use an artist for a composition they may have not created. For example, Edward Hopper died in 1967. So, he is unaware of laptops. Try this prompt now:
Businessman with a laptop computer by Edward Hopper.
Naturally not the greatest images by him, due to the lack of detail in the prompt. Plus, this raw version of Stable Diffusion is a bit wild with things like hands. But one can see the potential here to generate original images here with very little effort.
How to create a masterpiece
Go to art school for ten years. Just kidding! Ok, try studying the style of a few famous artists you admire. Then in Stable Diffusion use these names in combination in your prompt to form your own unique style. This coincidentally is also the reason artists are suing Stability.ai.4 Artists did not know, and therefore did not allow their images to be used in the training data. AI now makes it extremely easy for anyone to copy an artist style in Stable Diffusion. Interestingly to note, the artists are claiming copyright infringement, yet none of the artists’ images are stored in the training data and no new images are copies of the original. So, this lawsuit may take a while to sort out.
Personally, I’m pretty invested in integrity, so I generally like to use only artists that have passed away and always in combination. Apart from the artist’s name, also add in some extra detail into your prompt of how you want the image to look.
A good prompt may look like this.
Landscape image of a forest, atmospheric, hyper realistic, epic composition, landscape vista photography by Carr Clifton & Galen Rowell, Dustin Lefevre, painting by Ivan Shishkin, 16K resolution, trending on Artstation, DeviantArt, Flickr, detailed post processing, rendering by octane, unreal engine
Creating good prompts is an art form on its own and as with ChatGPT, the power is in the detail you provide to get the results you want.
Photorealistic ice cream Sundae in a bowl sitting on a rustic wooden table, coffee shop background, cinematic lighting, delicious, glistening, chocolate sauce, caramel sauce, swirl, drizzled, highly detailed, depth of field, Leica M6, Fujia Velvia, F1.4
I’ve put some old camera detail at the end of the last prompt. Interestingly, Stable Diffusion allows you to use highly rare and expensive cameras that you could never afford to own in your prompts.
How to detect if an image was created by AI
Last year I could give you a dozen telltale signs of an image generated by AI. Some of these signs can still be visible in some iterations of the software. For example, as mentioned before, hands are generally difficult for the software and often you would get deformations or extra digits showing. However, specialised trained AI image models are now available, usually for a fee, that can generate superb images with very little prompting. Have a look at the opening page at this website https://lexica.art/. All the images here were created by AI.
I think the only clear sign in the future that an image may have been created by a human instead of AI would be to see some iterative creative steps of the image, as a human would naturally do. Though it is also possible for AI to generate a sketched version of any image too, so this may also become a difficult thing to ascertain as well in the future.
1. I- Robot, by Isacc Asimov, (Movie 2004, 20th Century Fox, https://www.imdb.com/title/tt0343818/)
4. https://www.cbsnews.com/news/ai-stable-diffusion-stability-ai-lawsuit-artists-sue-image-generators/