The weird and wonderful world of AI-generated images for cycling
A few years ago we at CyclingTips wrote an article tackling the (slightly silly) question: Could AI be about to steal our jobs as cycling writers? The answer, of course, was “no” (at least not yet), but it was still fun to dive into the fascinating world of neural networks and AI-generated text.
The concept behind the technology is simple: give the AI engine a text prompt, and it will spit out a long passage of text for you, building on your prompt and drawing from the myriad of texts it was trained on. The results are often fascinating and more than a little entertaining.
Today we dive back into the world of neural networks with a look at AI-powered technologies imagery. Take note, professional cycling photographers.
Over the past two years, we have seen an explosion in the number of AI image generation platforms available online. You may have heard of DALL-E 2created by the OpenAI lab founded by Elon Musk, or its lite cousin Crayon (formerly DALL-E mini). Mid Road has also been causing a stir in recent months, especially in the fantasy art world, but these are just a few of the many options available.
These platforms all work much like the AI text generators we talked about last time. Under the hood of these sites and apps are neural networks that have been trained on huge libraries of existing images and their descriptions to understand what certain objects look like and how they interact with each other. When you feed these AI routines with a text prompt, they do their best to create an image that matches that prompt, based on the learning they have done.
Some of the results are truly magnificent.
After observing the growth of this space, it made us think: what is possible in bike with AI-generated images? What would these AI platforms create if we fed a bunch of cycling-related prompts? And would any of them be suitable for use on CyclingTips, for example?
And so we played around with a bunch of free tools: ensil, Crayon, Steady broadcast, Mid Roadand dream studio, feeding a bunch of different cycling-related keywords. The results, we’re sure you’ll agree, are fascinating, entertaining and, in some cases, more than horrifying.
We started with a simple prompt: “a cyclist going up a mountain”. Here is what we got from DreamStudio:
There’s also a lot to take from this DreamStudio effort:
Here are some attempts at Stable Diffusion:
And some from Enstil:
Next, we tried the word “platoon”.
Here’s how Enstil interpreted “platoon”:
Alright, let’s say you’re a bike designer and you’re looking for some futuristic ideas to help you with your next model. Maybe you look to AI for inspiration? Here’s what you’d get if you dropped “Specialized Full Suspension Gravel Bike” in DreamStudio, for example.
Speaking of a lot of hits…
Let’s take a look at what some AIs do with a “state-of-the-art bike in a showroom.” First, here is the stable broadcast:
Looks like we don’t need to send James or Dave to Eurobike next year.
Enstil also had some ideas:
Perhaps the showroom DreamStudio envisioned was a little too crowded.
Here’s Midjourney with the prompt “a showroom full of futuristic bikes”:
Changing gears, how about some close-up portraits of cyclists?
The practice of gravel is all the rage at the moment. Let’s see what the AI does with it.
Here is “A Cyclist Celebrating Victory in a Cycle Race”, first from the Enstil:
Here’s Stable Diffusion’s effort with the same prompt:
Speaking of faces, Craiyon seems to really struggle with this.
Here’s a Stable Diffusion series that went surprisingly well from the “black and white, filmstrip, pained face, cyclist, bicycle” prompt. Anyone looking for a cover for their new indie rock album?
Alright, let’s go through a bunch of additional images generated from a bunch of different prompts.
Alright, so a lot of this is pretty dumb and unusable. But there are ways to get the AI to create images that are a bit more useful. Using more feature-rich and well-developed AI platforms is one way (DALL-E 2 seems to be the gold standard but access is limited at the moment). Or you can try asking the AI to be more artistic rather than photorealistic.
Speaking of artistic, it’s possible to generate some really wonderful stuff with Midjourney. Check out these two stunning images generated from Leonardo da Vinci’s “sketch of a futuristic road bike” prompt:
Better yet: how about some Pixar portraits of some pilots?
These are just some really awesome images from Midjourney. Which brings us back to one of our original questions: is AI capable of generating images that are good enough to use on a professional website?
For most of the images above, the answer is a pretty decisive no (as fun as they are to watch). But in the case of more artistic or abstract creations? There is obvious potential there.
At CyclingTips, we often look for non-racing photos to help illustrate articles on various topics. It can be difficult to find a photo that talks about, for example, road safety or the mental health benefits of cycling. AI could clearly help in this regard, particularly if a more abstract or stylistic image is appropriate. As you can see above, Midjourney already allows you to create great content just for this purpose.
There is, of course, an ethical dilemma to be resolved here. Using AI to generate images like this directly takes jobs away from those who do this stuff for a living. It is no surprise that there are considerable controversy on technology for this reason. That and the fact that these platforms effectively generate images from different images that others have originally created – in itself an ethical gray area.
We are not going to solve these problems today. What we can say, however, is that race photographers need not worry. Example: Here is Enstil’s interpretation of Remco Evenepoel’s victory at the Vuelta a España.
And here is a preview of the final podium of the Vuelta 2022:
As far as racing footage goes, at least, we could stick to the pros for now.