I am using SwarmUI's default one - not tested custom ComfyUI workflows... Just a diffetent toolset. Inpaint in particular in SwarmUI is done like this (identical approach for Krea2 - just not using Flux specific settings, CFG od 1 and steps 8 to 20 - better 20; this LORA - for me works 1.0 weight, maybe 0.8, more than 1.2 starts making legs bare, 2.0 and more makes them nude no matter the prompt): https://youtu.be/09UsITSDJmA

And inpaint as it is implemented in my personal opinion is a bit unreliable especially for distilled versions such as Turbo - with CFG 1.0 I mean, can bring sometimes extra limbs or so, it's more of a fighting the tool than enjoying flawless cooperation (extra limbs, colour mismatch for skin is a a regular Tuesday for Inpaint). But maybe I am using it not the most optimal way, dunno - not a lot of good tutorials exist. Maybe I will write mine later on, based on my experience. That's why I mentioned that this model is highly experimental. Actually, for Krea2 specifically version I managed to get some descent results (as in main image) even using it as non-inpaint version, but as it was trained on a dataset having only part of body it's kinda a lottery - might work, might not. SDXL typically cuts the head and upper part of body in that situation, only smaller weights work but that defeats the purpose. So - tricky as of now, experimentation is encouraged.