AI image generators have taken the creative world by storm in the past year, unleashing a wave of stunning surreal images created by text prompts. But certain subjects can still be a challenge for their algorithms. And it seems that includes cycling.
The website CyclingTips had observed the furore around the new technology and decided to put it to the test, asking several AI models to generate images of cycling-related scenes. Close your eyes now if you want to avoid a horror show. Some of the results look like the aftermath of a particularly horrendous Tour de France pile up (to learn how the tech works, see our guide to how to use DALL-E 2).
CyclingTips was intrigued by the impressive AI-generated imagery that's been appearing since text-to-image generation took massive leaps forwards with the beta releases of diffusion models like DALL-E 2, Stable Diffusion and Midjourney. It had to try out the tech for itself.
"What would these AI platforms create if we fed in a bunch of cycling-related prompts? And would any of them be suitable for use on CyclingTips?" it wondered. It used several AI models, including Stable Diffusion and the free browser-based app Craiyon, and wrote text prompts asking for images of different aspects of cycling, from “a cyclist riding up a mountain” to “a show room full of bicycles”.
It found to the results to be "fascinating, entertaining, and, in some cases, more than a little horrifying." Yes, it seems that AI image generators struggle with cycling even more than they do cereal boxes.
The site was pleased with some of its results for “black and white, film still, pained face, cyclist, riding bicycle”, suggesting they could look great on the cover of an indie rock album. Stability AI's DreamStudio was less successful in rendering the prompt “Exhausted cyclist slumped over bicycle with sweat on their forehead with dark skies and lightning”. This poor guy is going to need some serious physio.
So what went wrong? One thing that might be causing problems is that many of the images generated are horizontal. Most AI image generators can only currently generate square images. Stable Diffusion allows different aspect ratios to be chosen, but, because of the images it's been trained on, it doesn't often seem to do a very good job with images that are wider than they are tall. It often fills the extra space with duplicates, giving you two of whatever you asked for.
Certain descriptions can generate strange results too. It seems some generators take "close up" to refer to the proximity between subjects in a composition rather than the type of shot. And AI doesn't know what looks right or realistic. It doesn't know how many wheels a bike should have, for example. Their training data no doubt included images in which the wheel of a bike was obscured behind the wheel of another or in which not all of a person's fingers were visible.
Some of the glitches and artefacts in AI-generated images can be corrected in editing software, but in this case, a lot of these are complete right offs just like the mangled bicycles depicted. To see more on how different AI image generators, see this comparison of the best AI image generators.