I decided to test an AI image tool on an actual project recently, and the results surprised me in more ways than one.
I had a diorama design I'd already built, and instead of starting over for Halloween, I took a screenshot of the pumpkin lantern designs and asked Gemini to keep the structure but reimagine the content for Halloween. The structure came back exactly the same... I wasn't sure it would. And the Halloween content it came up with was genuinely intriguing. It moved between different themes easily, more easily than I expected to see from a tool that had never touched a diorama before.
Some of what it generated was too detailed to actually cut. The images were intricate, sometimes beautifully so, but far more complex than a piece of material can hold together once you start removing sections. So I'd ask it to simplify. And that came with its own problem... when I asked it to pull back the detail, it didn't always take out the right things. Sometimes it stripped out details I actually liked along with the ones that had to go. Getting it to simplify precisely, keeping what worked and losing only what didn't, turned out to be its own back and forth. And it was frustrating in a specific way... once it changed the design, there was no going back to the original. Every version after that point had to be modified forward from wherever it had landed, not undone.
Then came the part where I actually had to build it.
AI has no real concept of negative and positive space. It has no sense that a piece gets cut out... that the detail it just designed is sitting exactly where the blade goes, and it's about to be lost to the workroom floor. It designs as if everything survives the cut, because as far as it's concerned, nothing gets removed.
So then came the real work. About an hour of point editing, piece by piece, fixing everything that assumed material would still be there once the cutting was done, and restoring some of the details it had simplified away. That hour is what turned an interesting image into something I could actually assemble.
I'll admit, this left me a little unsettled. If AI can generate something this good in seconds, what does that mean for the years I've put into getting good at this? Is this still mine to claim... or did I just clean up someone else's idea?
I asked Cynthia Solomon what she thought of the AI designs, and her answer stuck with me. She called it a total interactive behavior... the human has to be willing to experiment, and has to stay available the whole way through. AI can't do everything on its own.
Here's where I've landed. The idea generation was fast, and I'll give it credit for that. But it has no idea what a cutting blade actually removes, and it can't tell my favorite detail from a throwaway one when I ask it to simplify. It has no sense of what survives at scale, or of the hundred small judgment calls that separate a pretty picture from a buildable object. That part is still entirely mine. It's the part I've spent years building, and it's the part that made this diorama real.
Materials
I use Neenah 65 lb White Gold metallic cardstock from Office Depot for each panel and black 65 lb cardstock for the backing of each panel
Cut Files
You need an electronic cutting machine for cutting the files. The .Studio file is for the Silhouette machine and the SVG file for the Cricut machine.
Here is the .Studio file.
https://drive.google.com/file/d/1DmH_zWatJyMihhvL0dnhB-V0z2YCDQrN/view?usp=sharing
Here is the SVG. The file extends beyond the viewable area. Zoom out to see the entire file.
https://drive.google.com/file/d/1DU9QwVfwR_d4CjhACk16adrYD2dUQL3i/view?usp=sharing
Assemble The Pumpkin
Follow the assembly directions from the hexagonal vase project to construct the structure.
