Motifsnap

The status of AI art boom

It was often assumed that creative labor would be among the last to be mechanized. Some may rethink after 2022.

Last year, numerous sophisticated tools for generating art using AI just by putting in a few phrases were widely accessible in a matter of months. The quality of the graphics, pictures, and paintings that may be created in this manner has greatly increased. Some commercial artists are experimenting with the technology, but not all of them enjoy it, and stock picture agencies are planning to sell AI-generated photographs.

Because of this quick advancement, entrepreneurs have been rushing to develop goods and businesses based on AI picture generators. The technique is still being refined by researchers. WIRED recently had the opportunity to test one of the first AI systems capable of making video, built by Meta researchers. The films aren’t perfect, but when compared to examples from years of study leading up to 2022’s AI art boom, they present a visual chronology of a technology growing swiftly from lab experiment to commercial prototype.

Image 69
Courtesy of META: A video generation system created by researchers at Meta produced this clip of “Fireworks over Manhattan.”

The image-generation technology that has captured the interest of businesspeople and artists is based on decades of AI progress. Around ten years ago, researchers discovered that feeding neural networks vast volumes of photos with related labels allowed them to categorize previously unknown images with excellent accuracy. This is how Apple Photos and Google Photos can automatically categorize smartphone photos of dogs.

Image-making AI technologies reverse this image-labeling approach. Algorithms that have devoured massive amounts of photos and related information from the web may produce new images based on text input from a user. A “generative model,” which learns the features of a collection of data and can subsequently produce new data that statistically fits in with the original collection, lies at the heart of the system. This method may be used to create visuals, write text, produce music, or answer questions. The business potential of so-called generative AI has piqued the interest of technology investors.

Although generative models have been employed in statistics for decades, last year’s AI image-making boom stems from a 2014 discovery. That’s when Ian Goodfellow, a student at the University of Montreal at the time, devised generative adversarial networks, a novel approach to generative models (GANs).

Image 70
Courtesy of Ian Goodfellow: In 2014, an algorithm called a GAN generated these faces. The rightmost column shows real photos used to train the system.

GANS involves two neural networks (machine learning algorithms) competing against one other. The first attempts to produce something that matches a set of instances, while the second attempts to discern between actual and false examples. The fake detector forces the fake generator to improve over many rounds of competition. This approach produced basic representations of handwritten text, crudely drawn faces, and more complicated settings that looked like actual photographs.

Image 71
Courtesy of Alec Radford: In 2016, after digesting 3 million photos of real bedrooms, a GAN generated these rooms of its own.
Image 72
Courtesy of Phillip Isola/Alexei Efros: In 2017, a project called CycleGAN showed algorithms could remix visual components from different images.
Image 73
Courtesy of NVIDIA: Every one of these faces was generated by algorithms trained on 70,000 photos of real people.
Shopping cart close