What is 'generative art' and how does it work?
Using a Artificial Intelligence to create art is as simple as typing a few words
Computers have been using artificial intelligence, or AI, to create art for decades. However, with the exploding popularity of phone apps like Dream, Wonder and Starryai, it has never been easier to get started with “generative art.”
The one thing these apps have in common is that they allow users to enter a text prompt, i.e. a description of something, that the AI quickly interprets as an image in the artistic style that you specify.
Here’s an example of the user interface on Dream and the image it created after I typed “Smells like teen spirit” into the prompt box and chose the “spectral” artistic style.
The AI seems to have interpreted it as you might expect: young, ghostly figures walking through a creepy scene.
But what process did the AI go through to come up with this image?
When you enter a prompt, the AI reads it and tries to understand what you’re asking. The phrase “Smells like teen spirit” is a reference to a song by the grunge band Nirvana. While I may have been hoping that the AI would riff on the Nirvana reference and produce something that alludes to the song lyrics or even the music video, the AI does not have my frame of reference.
That’s an important thing to keep in mind when writing prompts. Try to write them as if you’re explaining something to an alien who has just arrived on earth today and who can understand your vocabulary but lacks context and meaning.
Here is the oversimplified explanation of how the AI processes your text prompt: The AI sees the key words – smells, teen, spirit – and starts scanning a vast data set of images, looking for images that might be tagged with those descriptions or may relate to the AI’s understanding of those descriptions. The AI then uses those images in the data set as reference to create a new image.
The actual process is a bit more complicated. Dream, like many other AI-art platforms, depends on two artificial neural networks that work together to process your text prompt. One of these networks scans through a data set of images, looking for images that best match the text descriptions in the prompt. The network then feeds that information to the second network, which generates an image that is similar to the ones the first network identified.
The two networks then pass the image back and forth hundreds of times, making refinements in a process that takes just a few seconds.
Obviously, there are myriad factors that can affect the image generation. One is the quality of the data set that is being used. What images does it contain? Any image that the AI creates is going to be a derivative of those images in some way. When you ask the AI to copy an artistic style, in our case “spectral,” the AI is looking at all the other images in its data set that have been identified as the spectral style.
Another important factor is how you construct your text prompt. There is not necessarily a wrong or right way to do it.
Being very specific helps the AI narrow down the possibilities, allowing it to potentially generate an image closer to what you had in mind. When I expand my text prompt to “Curt cobain singing the song smells like teen spirit” (even though I misspelled his first name), it takes the AI in a completely different direction.
But there’s also absolutely nothing wrong about being vague or using simple descriptions and letting the AI take you along for the ride.
While Dream is a popular app (Google recently named Dream its app of the year), there are other, more powerful AI-art platforms that give the user even greater control of how to shape the image.
Three of the best are Stable Diffusion, DALL-E 2 and Midjourney. My favorite is Midjourney because I’ve had the best results with it. However, people are also creating amazing images with Stable Diffusion, DALL-E 2 and many others.
One of the most exciting (or scary?) things is that the AI on these platforms is constantly learning and “training” itself based on the images it’s being asked to create.
The best thing to do is try several AI-art generators and see which ones meet your needs and budget. Some apps will let you do quite a bit with their free version. You’ll usually have to pay for greater processing power, speed and higher resolution images.
The cost for premium accounts can range from a few dollars a month to the $30 a month “standard” membership with Midjourney.
If you are trying AI-generated art for the first time, I’d recommend starting with Dream because it’s simple to use and allows you to create images (in limited styles) for free.
When you’re comfortable with Dream, check out Starryai on your phone to gain greater insight into how to create more advanced prompts.
Another great platform is NightCafe, which lets you use your web browser to create images with the Stable Diffusion and DALL-E 2 algorithms. NightCafe works on a credit system. You’ll get some starting credits to play around with and then can buy additional credits. You can also earn credits just by logging in each day and claiming them.
After you’ve tried out some of these options, check out out Midjourney to see how it compares. I’d recommend waiting to try Midjourney until you have a good idea how to use advanced prompts so that you can get the most out of its very limited free trial.
But consider yourself warned – once you start with AI-generated art, it can be highly addictive. I’d love to hear in the comments what your favorite AI-generator is and why.