Do you think this will change and get better? Would it help if the AI was capable of interactively refining the inputs (I guess by asking questions instead of making assumptions)?
In theory, given enough machine power, I could see a branching, interactive interface. Like if there are multiple major hills in the latent space ("Tempest-> a storm cloud" vs. "Tempest -> a cartoon pony") the system could identify them all and generate a set of images with different biases towards each hill. Then the user can pick which direction to climb.
The current system on high-end hardware takes about a minute before results for one image even begin to be recognizable. And, about 3 minutes before you can be somewhat confident how it's generally going to turn out.
The system is garbage in garbage out. So, feeding it random noise inputs will give noise results. The human factor is “prompt engineering” the inputs to try to coax the elephant in the desired direction. And, in evaluating the outputs.
Sometimes it feels like asking a child to draw something for you. I spent some time trying to get “storm clouds made of lava” but the child insisted “No, dad! Lava stays on the ground. Like THIS!” https://www.reddit.com/r/bigsleep/comments/o033x1/playing_wi...
Similarly, while trying to make a “tempest in a teapot”, instead of a storm, the child was a fan of the My Little Pony character Tempest Shadow. https://www.reddit.com/r/bigsleep/comments/o04sin/trying_to_...