Road bike AI: Does artificial intelligence realistically depict road cyclists?

One thing is clear: in a few years, the internet will be flooded with computer-generated "photos". The biggest fear is that we will no longer be able to distinguish between fakes and real photos. Even today, programmes such as Midjourney can create deceptively real images of racing cyclists in just a few seconds. In this article, we want to look at what ideas AI (artificial intelligence) has about racing bikes, racing cyclists, bike races and other topics related to narrow tyres and how to get good pictures. Let's start with a simple task: "Take a realistic photo of a typical road cyclist on a country road".

The app displays four variants and you can see immediately that the AI mixes road bike and motorbike. This points to the ambiguity of the term "bike" in English. So let's try a more detailed description: "Cyclist road bike small road from front realistic photo daylight"

From the front, the AI didn't understand and there are only two racing bikes in the selection. If you create a lot of pictures with Midjourney, you often get a result that looks like sunrise. Perhaps there were a lot of early riser photos in the training data for the racing bike AI? Out of interest, let's try to see if the result in English ("cyclist roadbike small road friontal view realistic photo normal daylight") turns out better.

Racing bike KI Midjourney is better in English

The result is much better and only racing bikes can be seen. The style alternates between photorealism and art - this is a question of the setting or the formulated order to the AI. Because we are using the web version of Midjourney, we are limited in terms of settings. What is striking is the sometimes poor road. AI needs clear instructions - that's why we create a prompt (that's what the commands to ChatGpt & Co. are called) that gives a lot of details: "Group of cyclists on small paved road in Austria in the Alps. background high mountains normal daylight open landscape. Realistic photo Nikon style."

At first glance, the result of the road bike AI is not bad at all - that's why we had the four variants enlarged individually. The mountains look impressive and the narrow road appears to be car-free - perfect for road cycling. It is noticeable that many road cyclists ride with rucksacks and that men dominate. But if you take a closer look, you will discover a few mistakes:

With Midjourney, the computing power is concentrated on the foreground - this is where the images are detailed. The background is often blurred or contains errors. Enlarging the images reveals a few anomalies: missing or crooked legs, bizarre pedals and strange panniers. In the small preview, however, Midjourney's pictures are quite good. If you look closely, you will realise that they are not real photos. The central question is: how does the software come up with the pictures?

Racing bike AI: how does Midjourney work?

Let AI provide the answer: this is what LibreChat says about how the software works:

Programmes such as Midjourney use a process called Generative Adversarial Networks (GANs) to generate artificial images. GANs are a type of artificial intelligence that aim to generate data that is indistinguishable from real data. GANs consist of two parts: a generator and a discriminator. The generator tries to produce data that looks real, while the discriminator tries to distinguish between the data produced by the generator and real data. During training, the generator gets better and better at generating data that looks like real data, while the discriminator develops better and better skills to distinguish real from fake data. In the case of Midjourney, the "data" in this context means images. The AI is trained with many real images, such as landscapes or portraits. The generator then learns to generate images that look like real images, while the discriminator learns to distinguish them from real images. Ultimately, you end up with an AI that can generate new, convincing images of, for example, landscapes or faces that it has never seen before. It is important to note that although the generated images may give the impression of representing real subjects, they actually originate entirely from the AI's "imagination" and do not correspond to any real template.

A look at the proportions and details still reveals weaknesses. How the prompt is formulated is also crucial. If you take the example above, you would try to recreate the unrealistic street again.

Generative AI is improving at a rapid pace. And it is to be expected that specialisation will take place in certain areas - e.g. images of people in nature. In a few years, it will probably be difficult to distinguish between photos and AI-generated images. Today, a glance at the details is enough to find many errors. The Leonardo AI image generator is not one of the best when it comes to racing bikes. The centre stripe moves to the edge, the left-hand racing bike has no saddle and the entire frame of the right-hand bike is crooked. AI finds it particularly difficult to generate transitions - this is clearly visible on the right leg of the rider in the white jersey.

Conclusion: fortunately, we at TOUR can continue to rely on photos by professional photographers!