Generating trash pandas with Copilot

I was chatting with someone about some picture and I noted that maybe we could find something appropriate.

I prompted DuckDuckGo with “raccoon trash panda digging in garbage can”. It gave me a whole grid of pictures, of which I quickly picked the following as my favorites.

Trash panda relaxing in a garbage can as if were a hot tubTrash panda peeking out of garbage canTrash panda living its best life in a trash canTrash panda head down up to its hips in a trash can.jpg

My interlocutor had never really used any of the LLM-based machines before, so I gave GitHub Copilot a whirl. I prompted GitHub Copilot: “Show me a cartoon of a raccoon digging in a trash can with its butt in the air”. It gave me the following:

Raccoon in a flowerpot with a monkey's assRaccoon in a trash can with dangling feet

The second one is not bad at all, but the first one! What the heck? Is this what Copilot thinks I want when I ask for a cartoon of an animal? Maybe most people using these things are there to sexualize animals?

At any rate, I figured it should cover it up so I can use the graphic in SFW situations.

Prompt: “Put some orange swim shorts on him”

Raccoon in a trash can with an orange swimsuit and an offset tail

As I’ve experienced every other time I’ve tried generating images, there is just no way to get it to build on the previous image. You can’t include it as a template. A human artist would have given me the first two images, but with an orange swimsuit. Copilot gives me a completely different graphic of a raccoon towering over a garbage can, but with orange swim shorts, as requested.

Its tail is kind of just pasted on. Why? Because Copilot has no idea what a tail is. There’s just a high probability that there’s something stripy at the other end of a raccoon from its face.

The images I got just on the first page of DuckDuckGo are all way better. That took me seconds. Generating images takes a lot longer and never really gets to the desired result.