I've been playing around by attempting to make Stable Diffusion "hallucinate" in creatively useful ways. For this LoRA I used a large number of digital collage/abstract works by Greg Ellison, and tried to train it to separate the abstract elements from the stylistic as an experiment. The resulting model doesn't produce generations that look like Ellison's work much at all and wasn't really meant to, but produces generations with wildly varying stylistic effects. It must be played with.

I'm (mis?)using "concretion" as in a kind of de-abstraction. Perhaps "concretization" would've been a more accurate term, but I'm frankly not sure what to call what this does. I'd probably rename it if the name wasn't already embedded in the LoRA.