For my part, my production process goes through 2 models, a turbo on 5 to 8 steps, which allows me to see if my scenario is well understood, this model makes a lot of errors in interpretation, but it gives an idea and especially in 0.4 megapixel it allows me to know if I am going the wrong way or not. then I switch to a 20-step fl2va model for the final, it's much longer, but the quality is really better even if aberrations can still be generated of course.
I have a real challenge with the weights of the Lora I use, and they are not all compatible with the distilled models, but overall I find that just with Lora for both male and female sexes, the minimax models work incredibly well, much better than Wan and LTX, in terms of quality and understanding of the scenario. I sometimes even have the impression that it understands the context by putting the right tone in the texts that I generate, these models are crazy!