I read somewhere that for making concept and style models like this, less is more, and to try to work with 10-20 carefully curated and carefully tagged images. So I tried taking some of my big favorites as only tagging the small handful for this training. And honestly? I got a lot more out of this than I expected.
So I'm thinking I'll do separate sets of "looner" models based on specific poses or actions involving balloons pulled from my hoard. This is specifically "balloon straddling", not "sitting" or "riding" or hugging or any other variation.
I find best results with 20-30 steps, Euler Ancestral, CFG 6.5 (really, anything from 4 to 7 seems good, but I've had the most consistency with 6.5.).
I've got like 100 generations today that I'm pretty happy with, I'll try to add some extra posts with some of those shortly after I get the model up.