>>3027645This is a good example of the randomness of Stable Diffusion predictive algorithms. Technically, pretty much everything
>>3027634 said is good advice. However, using tokens can have unexpected effects because of the random noise and possibly unexpected model token associations. Things like "four fingers, one thumb" aren't tags that models should know, but though base SDXL knowledge or even random chance they could still help gens be more consistent.
Ultimately you should do whatever gets you the results you want, even if it doesn't make sense.
