>>8667296
>Can you explain what these steps and
invert mask does exactly what it says. It makes a mask of what you didn't select (inpaint). Image Composite Masked takes the original, before inpainting image, and the output of your ksampler->vae decode, and blends them together, so that the original image's pixels are used in the areas you didn't explicitly inpaint
>this step is for?
It takes your original image and applies some magic I do not understand to your clip conditioning that makes the inpainting respect the original image's style better even if you don't prompt for it
if you looked at my prompt you would see I didn't prompt for any artist tag and yet my edit somewhat resembled the original
here's what happens if you use my workflow without the node and with that same basic, almost empty prompt:
https://files.catbox.moe/ac1p3y.png
There may be times when it's better to inpaint without that conditioning node if you need more agressive edits, but overall it tends to stay in most of my usages.
>(as well as the rescale and differential diffusion nodes)
https://differential-diffusion.github.io/
if you need the full explanation
personally it makes my head hurt and I don't care, I'm grug, I see something work in trial and error, I keep it
RescaleCFG to be vulgar and unscientific basically fixes the screwed cfg of vpred that makes anything over than 5 burned and also makes the lighting behave more like eps models so that you don't constantly get very dark scenes if you don't prompt for light/sun etc
you don't need it with WAI or other eps models
it's only for vpred, and only really useful for base vpred with low strength loras
most of the vpred shit mixes pretty much convert it to behave like an eps model