nyanko7

/

flux-dev-de-distill

like 159

what you mean? on inference this model works best with cfg scale of 3.5

CFG = 1 is actually a disaster with this model and pointless

Yes off course, it uses real cfg scale, it works with higher than 1 cfg 3.5 is the default you can try other values, but trained loras on this model work on cfg 1 on inference in conjunction with the regular flux-dev model or hyper-flux-dev

See how you like these, I used 2 LyingSigmaSampler to reduce detail some. All prompts from a Tom's Hardware article. I didn't adjust them any, and I just used 1492 for my seed.