>>2861288a more fair comparison
>>2861289IKR
I can't find the paper now, but it was about modifying the model training to align the inference with the low steps used commonly instead of assuming the same same step count as training
Altenatively, there is also this, which is training free
https://arxiv.org/abs/2106.03802>>2861296Both AWA and Olympus v2 works okay with Hyper-SDXL, you can have results in 10 steps when combined with AYS
Olympus v2 works fine on CFG 1.0, but AWA needs higher CFG (try 4.0)
