Usually when generating anything, whether images or otherwise, I've seen a lot of safety checks that models have.
Paid services are especially good at that. There are many things you can't make them do or generate, like violent or edgy content. Local models tend to champion freedom in that regard, which is why a lot of people use them, but this time I ran into a strange problem.
It was the problem of the closed curtains being opened once morning came.
I clearly wrote a negative prompt saying "don't open the curtains even when morning comes!" but it didn't work, and even when I gave a clear action instruction to not open the curtains and had it do something else instead, the curtains would still open automatically. It was truly maddening (huh?)
Whether I gave a simple scene instruction or wrote out a complicated prompt describing everything in detail, the curtains kept opening as if possessed... I spent a whole day on this issue. In the end, I only managed to catch the haunted curtains after inserting a reference image of the room scene itself and locking it in place. It ended up being the most time-consuming production yet.
*This song is from the Epik High song festival, where Tablo expressed through song that his wife Kang Hye-jung watches the animation she used to watch as a kid, alone, whenever things get hard. If you go to the Epik High channel and watch around the 16-minute mark of Part 2 of the song festival, you'll see the story.
I had a similar problem with something else — while making Lisa's SaWaDiKa, at a certain point the output would come out looking almost photorealistic instead of a 2D image. In particular, once "Lisa + fashion show" were combined, it would transform into a photorealistic output wearing a fashion show outfit that I never even specified. And depending on the outfit, some similar outfits would also immediately get converted into photorealistic output. Still, this was a problem I eventually managed to solve with a strong enough prompt.
I actually ran into this issue for the first time back with Bigbang, but at the time I moved past it quickly using a character reference, so I didn't really notice it — it turned out that pre-trained data of celebrities could pop out saved training images just by calling their name.
There's also material someone tested on this.
https://huggingface.co/datasets/malcolmrey/various/blob/main/h3-center/known-characters/INDEX.md
These are cases where specific scenes can be generated just by calling a name.
So ever since then, when depicting a real singer, instead of using their actual name, I came to an agreement with the AI to simply describe them as "picture 1" (..). So if you're working on a production and no matter how much you explain and fuss with the prompt it just won't work, quickly give up, revise the scene content, or move on to another method fast. There are no exceptions to the rules, not even for my own wife