Based on a more realistic Flux krea dev version training, it combines generalization and control, with good performance between weights of 0.6-0.8.