I took up David Blacker’s prompt to create AI images of armoured trishaws. It turned out to be a bigger challenge than I thought. Multiple Generative AI (GenAI) platforms had a hard time creating an image that matched the armoured trishaw I look I had in mind.
The first GenAI I tried was mid-journey (last year). It kept defaulting to 4-wheeled vehicles. Had the same issue with Google’s Bard and Pixlr.
The images in this post are more recent attempts combining Gemini – Bard’) and Runway. Even that ultimately required some manual editing.
The resulting GenAI platform images have a surreal, dreamlike element. Visual renderings of image elements that attempt to mimic photographs are common. The result looks like fragments of images or smeared details.
Armoured trishaws are not a standard theme in anyone’s training data. This underscores that generative AI is a statistical sorting system at its core. Things with common data get precedence over the less common ones.
The hardest part of the prompt process was getting the AI to accept the armoured and three-wheeled part of the prompt. Across all the platforms I used, I kept getting images of four-wheeled jeep-like vehicles. Adobe’s FireFly kept ignoring that a three-wheeled vehicle had three wheels. Stable Diffusion’s free version was similarly unhelpful. The most successful prompt on RunWay was a variation of
Armoured tuk-tuk with all openings covered by metal, similar to a tank or armoured car.
However, the image was based on another image generated using Gemini. I used RunWay’s selection tool to replace parts of the Gemini-generated image gradually.
Getting the AI to cover the driver-side openings was a struggle. Adding armour on the windscreen was even harder. It took multiple turns with RunWay’s brush isolation tool to replace the windscreen with “armour like the front of a tank”. Ultimately, I ran out of free prompts, trying to get the wheels covered.
RunWay could not figure out how to render the front wheel with a covered housing. It was hard enough to make the back wheels thicker. In the end, I had to manually combine two images using GIMP with a quick layer masking job. So, this attempt had a manual step.
My conclusion is that most AI systems know what a Tuk Tuk is. Yet they have a hard time envisioning a fully enclosed Tuk Tuk—let alone one enclosed in armour.
It’s tempting to rant about how useless GenAI is at this stage with niche prompts. However, I didn’t spend much time keeping track of the prompt engineering aspect of the process. The prompt structure for each platform was different. Gemini worked with full sentences, while others, like RunWay, ran on terse phrases.
Photos of modified trishaws designs are more imaginative and claustrophobic.
What’s clear is a lack of past data to inform the AI.
Yes, armoured three wheelers did exist. Belarus has actually created a military tricycle. The Italians had the hardest core design. However they were a brief burst of innovation that reached its natural dead end – at least in our reality.
Another constraint are the prompts. Either my prompts were too brief. Or their details muddied the result. Most likely it’s a combination of both.
The next step in my efforts is using the Lambda Image mixer. I’m confident that pictures being worth a 1000 words will result in effective prompts.
A 3D human rendition captures what I had in mind. These samples are beyond what I could cook up. This one is certainly far from anything I visualised.
Enough theory.
Cerno 