I feel like 'AI Whisperer' will be a job title within the next few months if it isn't already. Having played around with Stable Diffusion (on a low powered machine) I've found that choosing the right terms and aesthetics in your prompt are huge influences between good and weird outcomes.
You're right on. I noticed that Stable Diffusion does well with some kinds of prompts, such as "in the style of <painter>", or with outdoor scenery, or with mythology and illustration. Yet any images of real people are massively distorted, as if Stable Diffusion doesn't understand what a human face should look like.
It seems that a common "pipeline" is to run the outputs of SD through GFPGAN to correct the faces. Although it can't help with the number of limbs or fingers.
For some unbeknown reason, a telephone company installed three new phone boxes on the high street in my town last month. Two are no more than 100 yards from each other. My cynical theory is that they are being used to track footfall using wifi connections (but I have nothing to prove that). Very odd considering most councils are aiming to reduce the amount of 'street furniture'.
