@_Envy_ Not at first, I first used the de-Distilled model following some advice from a Youtube tutorial and it was terrible xD Then lastly a friend found me some settings on Reddit with previous model+training adapter v2, it's a lot better but still not great. Maybe putting the learning rate from 0,0001 to 0,00025 also made a difference.
Anyway I think that Lodoss style is very hard for those recent models trained for realism ^^; Or I'm just doing it wrong !
The results are very strange to analyse like on the same training one image can look completely flat color with no detail while another prompt can give completely different result with too much details.
I might try again with simplified captions, I don't know since Flux I'm basically walking in the dark here xD ok, training number 5 ! xDD