Ending the Season Images

In October I wrote a post about a little experiment I implemented. An experiement I wanted to build since I was a teen. Now I’m removing this experiment again.

To recap in short, back in the days when I started to create websites, I was fascinated by the idea that the design of a website (or at least a header image) depicts its own little world, with its own simulated weather and urban scenes and people living there. So I did it, with AI, which renewed the header images every 6 hours.

The magic wore off kinda quickly. It was neat to see how it added little scenes around halloween, christmas and new years; but it was also predictable that it will create those scenes. Those special scenes weren’t actually special anymore. And over time, it felt like uninteresting noise. If an artist had created them, it would’ve been much more interesting. Even though it wasn’t 100% AI slop because it had an interesting twist to it, it was still: just AI. People don’t just care about the art, they care about the artist.

I never wanted to keep those images here forever, and I’m happy this idea is now out of my head.

Here are a few oopsies that happened during the experiment:

• To simulate the weather, the AI is prompted with the weather it generated for the previous days, to create a new weather for “today”. It then uses this weather to write a new prompt for the actual image. But it started to write small scenes into the weather report itself, so it continued to reuse this same scene, and for a while all images were in a town with cobblestone streets and a church and a riverside.
• The images need a vignette, which was added to the image in a second step: it first creates the image without the vignette, and then uses this image to generate another image with the vignette. I thought that OpenAI will automatically detect the image aspect ratio to generate a new image with the same ratio. But that was wrong, it started to create square images. So for a while, the vignetted images were weirdly cut off and too small.
• During the React2Shell situation, the image generation halted because I didn’t redeploy the cron. While I updated the Next.js version in my monorepo, there were no changes to the cron (because it doesn’t use Next.js) and the deployment still had the vulnerable version of Next.js installed which we didn’t allow to run on Railway anymore.
• The AI was prompted to create scenes in a german city. But it did not create fireworks at midnight 2025 → 2026, and no streets littered with fireworks, which is just unrealistic af.