Krea2 didn't seem to understand the concept, and since @Topplok2 had already collected the training data (cheers pal) I thought why not.Minimal effort on my part, using civit's auto-captioning which I'm sure isn't ideal... it works just fine but maybe I'll have another go at some point.Preview images were generated at 1 strength (or during training) and without other loras.

Files with SHA256: e36c1340325fb50a85768c00a032727d9ee48c0a93cc34f6f62fa4758e0f3edc