Ask HN: Why do generative AI fail ignore parts of the reqeusts?
They all tend to focus on parts of the text supplied as a description and leaving other parts ignored.
"An emperor penguin sitting on an ice throne wearing a flower in her crown playing Nintendo"
Will give me a penguin, with some ice cubes, sometimes a flower, never anything to do with gaming. A longer description tends to leave more objects out of the result.
I understand not getting the objects right, but I dont understand leaving objects out entirely. Yet it seem to occur on many different text to image platforms with different models.
Being completely ignorant of how these things work, I would have expected part of the process early would be to separate out all objects in the instructions given to feed them for rendering.
The parsing might even happen prior to the text to image prompt sending the instructions down to the model.
The same is true when I ask for programming scripts as well. It system often just leaves out things.
It would be helpful with: "I am sorry, I dont understand this part", or "maximum complexity reached, some instructions left out"
Hopefully someone explain the basic process so somebody most ignorant about how the systems work.