AI has become an integral part of my production workflow, both during concept development and production. Despite its impressive capabilities, achieving a consistent visual language still requires a considerable amount of trial and error. The challenge is not simply generating compelling individual images, but creating something that feels coherent rather than like a random sequence of prompts.
Captain Bavaria is a case study in which I explored different AI tools and workflows, not as isolated experiments, but as part of a continuous project with real creative and technical challenges.
The scenario is deliberately absurd, allowing me to test these tools at their limits: How well does AI handle graphic design when developing a complex logo? What is the most effective workflow for turning generated concepts into usable 3D assets and integrating them with AI-generated footage? And how do you tell the iconic story of a friendship between a man and a hyper-intelligent Alpine chamois?
Captain Bavaria: Logo Development
Every project starts with an initial idea from which everything else develops. In this case, it was a Bavarian superhero with the rather self-explanatory name Captain Bavaria, loosely inspired by Captain America.
As a small twist, I interpreted "Captain" not as a military rank, but quite literally as a ship captain, simply because I liked the idea of a Bavarian seafarer.
A superhero needs a strong logo: a visual shorthand that communicates both character and concept. For Captain Bavaria, I wanted to combine Bavarian and maritime elements without making the design feel overloaded.
I experimented with different prompts in ChatGPT and Google Gemini, continuously refining them and passing results between models to reinterpret individual elements from different perspectives. In total, I generated around 50 variations, with varying degrees of success and some rather curious results.
As is often the case with AI-generated design, no single result was quite right. I therefore combined the strongest elements from my two favorite versions and refined the final design in Photoshop.
S.S. Bavaria: Concept Design
A captain needs a ship. In this case, a houseboat on Lake Starnberg. I envisioned an old-fashioned steamer converted into a houseboat, surrounded by the warm atmosphere of the setting evening sun.
Eventually, I arrived at a design that captured the character I had in mind. Using Photoshop's Generative Expand feature, I was able to derive the additional environmental shots I needed from this single key image while maintaining a consistent visual language.
I then used projection mapping to turn the generated image into a simple 3D environment, allowing me to use it as a virtual set inside Cinema 4D.
A Man and His Chamois: Image Generator Tests
Another objective of the project was to compare different image-generation models within the same visual context. Since every superhero needs a sidekick, I decided to establish this relationship through a series of photographs. Generating two human characters has become relatively straightforward, so I deliberately chose an animal companion instead. For Captain Bavaria, an Alpine chamois seemed like the obvious choice.
What surprised me most was how far ChatGPT occasionally went with its interpretations, generating not only visual details but unsolicited pieces of world-building as well. For example, the entire prompt for the following image was simply: "Man and chamois receive a diploma." Neither "Innsbruck" nor "Gamsbert" had ever been established anywhere in the project.
I still don't know what Alpine Excellence is, but apparently I have a degree in it.
AI Robots: From 2D Concept to 3D Asset
Every superhero also needs an antagonist. Given the project's overarching AI theme, robots were an obvious choice. My first instinct was the familiar image of an intimidating, Terminator-like machine, so I initially prompted designs in that direction.
Then I reconsidered the premise: If I were a highly intelligent AI trying to take over humanity, I probably wouldn't design my robots to look terrifying. I would make them as approachable as possible, increasing public acceptance and minimizing resistance. So I pushed the design in the opposite direction: maximum friendliness.
The final concept is cute and approachable without becoming entirely toy-like. From this 2D concept, I generated a basic 3D model using Meshy and then refined the asset further in Cinema 4D.
To animate the character, the model first needed to be rigged with a digital skeleton. I used Mixamo to automate this step. Mixamo also provides a large library of motion-capture animations that can be applied directly to the rigged character and then imported into Cinema 4D. This significantly reduced the amount of manual animation required and allowed me to focus on integration, timing and the final motion design.
