>>2758254I only have a layman's understanding of how it works, but its concept of "denim shorts" is just an abstraction based on everything fed into it with that tag. It's not like specific seeds will draw out a specific image. It has a concept of denim shorts based on the sum of its input. The limitation lies in the granularity of the tags it was trained with. If it was trained on a lot of shorts and only a few with frayed edges, it will have an overwhelming propensity to make them without frayed edges.
This is why a lora is so helpful. You are refining that training and feeding a better dataset in, so now it has a better concept of "daisy dukes".
On an unrelated note, here's a Lum from AOM3 A1.
