These AI Generated Scenes From The Great British Baking Show May Give You Nightmares


A far cry from the usual scenes in the baking tent. Janelle Shane/ AI Weirdness

The Great British Baking Show (actually known as The Great British Bake Off, or GBBO, in the UK) is as wholesome as TV shows get. Inside the infamous tent, contestants whip up baking delights, from bread lions, to sandwich cakes, whilst cheering each other on from their workstations. Therefore, more scenes from this beloved baking show could only be a good thing, right? Well, as research scientist Janelle Shane discovered, artificial intelligence (AI) had other ideas.

Shane trained a neural net, a set of algorithms designed to recognize patterns, with a selection of 55,000 screenshots from the show. Her hope was that the software would generate images in the style of GBBO, but the results were not as uplifting as she’d hoped.

Shane described the results as "less than cozy." Janelle Shane/ AI Weirdness

For starters, all the faces of the contestants were erased. A few iterations later and human faces began to reform but never quite back to the sweet-smiling expressions you expect to see in the tent. Shane explained in a post on her blog, AI Weirdness, that the system was overloaded with the variety of images she had fed it, which caused it to produce these uncomfortable scenes.

The neural net latched onto the bread texture (although it could possibly be human fingers), and filled the scenes with it. Janelle Shane/ AI Weirdness

StyleGAN2, the image-generating neural net Shane had used, is very good at producing human faces, Shane wrote, but only when the face is the only feature of the image. Even then the face has to be uniformly centered, otherwise the neural network struggles. Unsurprisingly, the GBBO is more than just faces; there are the bakes, of course, and the bakers' bodies and hands, the surroundings of the tent, and the occasional squirrel. The inclusion of these in the neural net’s training data resulted in scenes not dissimilar to a horror show.

"I fully expected problems," Shane told IFLScience, "there are so many different kinds of subjects and camera angles in the GBBO images that I guessed it wouldn't be able to learn them all. Even when I added 5x more data, it didn't really improve - the neural net just wasn't smart enough."

Your guess is as good as mine as to what this image is ...Janelle Shane/ AI Weirdness

Contestants had bread as flesh, their glasses appeared on culinary bakes (possibly…it’s kind of hard to tell), but amongst the chaos, the Union Jack bunting looked pretty normal. That, explained Shane, is because neural networks are good at producing patterns.


“Excessive repeating patterns is one of the hallmarks of neural net-generated images,” Shane wrote in her blog. “Even when the repetitiveness is more subtle, it still tends to be there, and it’s one of the ways you can detect AI-generated images.”

In Shane’s images, there is a whole lot of repetition, particularly where there isn’t normally any, like in human faces and bodies. Janelle Shane/ AI Weirdness

But what about the focal point of the show, the bakes themselves? They suffered less of a butchering than their creators did, including “a voidcake, floating dough, and terror blueberry,” as Shane described them. The repetition appeared less strange in baked goods than on humans, with some results not dissimilar to some dishes served up on the celebrity version of the show.

"Terror Blueberry Pie" anyone? Janelle Shane/ AI Weirdness

 Shane told IFLScience that there are ways to improve the process: "The best way to help it would be to divide up the problem into smaller bits that are easier for it to handle. Just closeups of the bakers' faces, for example. Or just finished cakes. I might try this sometime and see if it does better!"

If you would like a go at using artificial intelligence to ruin something you hold dear to your heart, Shane has directed people towards You just need several hundred pictures of your chosen subject, such as your cat – which I definitely don’t already have…


[H/T: Interesting Engineering]