Goseong Dinosaur Museum

I'd love to tell you I chose to test an AI-native pipeline on a real client project because I'm brave. I didn't. The company decided.

Goseong was a renewal project, and there was far too much to make in far too little time: twelve to fifteen videos, some of them multi-wall installations. The calendar said nine months, but four of those went to planning and client feedback, so if you're doing the math at home, that leaves about five months of actually making things. Goseong was the perfect project to go full AI, and also the kind of perfect you don't get to turn down.

So here is what “AI-native” turned out to mean. All of the content was generated with AI. Most of the iteration and fine-tuning was done by hand, because AI got us to the wall and people got it to look right. A few things, like the assets for the casual game content, were made by full automation. And underneath all of it sat a kind of work that doesn't look like AI at all: a dinosaur museum comes with homework. A species list, sorted by family and by period. Character sheet images sorted the same way, plus their styles. Environment data, sorted by period. Every time one of the experts weighed in, all of it had to be updated.

With that machine running, we went into our first review with three pieces: Dino Rising, Dino Stage Alive and Lake of Abundance.

About ten people were in the room, including a couple of city council members, and the mayor was one of them. The videos I showed were close to previz, meant to give a glimpse of how the final would look. The mayor took them for final, or at least close to it. He didn't like them. He said they all looked similar.

I should admit something: I liked the work. I liked it right up until the mayor disagreed. And I wasn't in a position to say anything after that.

We started from scratch.

I gathered the team. We had a lively moaning session, and it lasted a while. Then we started talking about solutions, and what we landed on was this: stop making three videos in one look, and make three different films, each with its own design direction. Before committing to any of them, we ran a lot of AI video tests inside that direction. Dino Rising became cinematic realism with a dynamic style. Dino Stage Alive became hyper-realism, almost a nature documentary. Lake of Abundance went to fantasy and surrealism.

Two of the three weren't single screens, which is where a lot of my time went. Dino Stage Alive has three screens, and I made the three videos separately, so getting their timing to match turned out to be one of the hardest tasks in the whole project. Lake of Abundance is an immersive installation in a square room, with three wall screens and a floor. I upscaled the generated videos four times and sliced them to fill all four surfaces, and for a couple of scenes I made a separate floor video, matched its colour and style to the walls, and combined it.

The second review worked. Everyone was satisfied, and the members of the city council were busy taking pictures in front of the videos.

If I could give you only one piece of advice before you try this yourself, it would be this: if you give AI full control from the beginning, it will give you dumb results. If you give it very specific, clear direction from the beginning, it will come back with unexpected and amazing ones. Everything else in this post is a footnote to that.