An AI-generated image of Angel Reese riding a unicycle
AI image generators struggle with unicycles. When prompted, Stable Diffusion 2 made this image of basketball player Angel Reese riding a "unicycle."
  • AI-generated art is becoming more popular with tools like Stable Diffusion and DALL-E 2.
  • But the technology has still struggled to perfect certain images.
  • The latest thorn in AI art generation's side? Unicycles.

Last month, a highly realistic image of Pope Francis wearing a luxury puffer coat set the internet ablaze, with a number of people believing that the Pope had a penchant for Balenciaga.

It turned out the image was made using the AI text-to-image program Midjourney, ringing the alarm bells surrounding the growing capabilities of AI image generators and the resulting possibility for the spread of misinformation.

But while AI art generators can depict a very true-to-life Pope, they still struggle with other images — including, it seems, unicycles.

Luke Bailey, the head of digital at the British publication The i Paper, recently wrote a Twitter thread about AI's difficulty depicting unicycles.

Testing various unicycle-focused prompts on text-to-image generator Stable Diffusion 2 revealed that the program was not able to properly portray the one-wheeled vehicle.

An AI-generated image of Shohei Ohtani riding a unicycle
When prompted, here's the image Stable Diffusion 2 made of baseball player Shohei Ohtani riding a unicycle.

The technology not only seemed to have trouble depicting unicycles themselves, but also humans' interactions with them. The generator seemed to have no idea where to place the human riders and spit out images that, in some instances, appeared to be more similar to bicycles. 

An AI-generated image of Mike Bloomberg riding a unicycle
When prompted, here's the image Stable Diffusion 2 made of Mike Bloomberg riding a unicycle within a city.

That said, AI may soon make strides in its depictions of unicycle.

Human hands had previously given AI image generators trouble.

Part of the reason for the difficulty is that these programs are trained on images sourced from across the internet: Faces, not hands, are usually the focus of pictures.

Speaking to The New Yorker, Alex Champandard, cofounder of AI firm Creative.ai, pointed out that these generative models would be better at portraying human extremities if most of the source imagery was "one hundred percent hands."

But a recent update from Midjourney has appeared to curve this weakness, per The Washington Post.

It is plausible to imagine that as more money flows into the industry and AI continues to make advancements, more of the kinks will be worked out — including unicycles.

Read the original article on Business Insider