VPred models are trained with V-Parameterization so they will only work on VPred models like NOOB-AI-XL-vpred or Obsession, etc.
This is a DoRA so make sure to have your stuff somewhat updated
Remember to check out the artist.
Will remove if requested
NoobAI:
For VPred version stick to NoobAI-vpred. It drastically improves style strength compared to Obsession. r3mix can somewhat maintain the style. You can use r3mix to generate the initial image, and then switch to NoobAI-vpred in Hires.fix pass.
Turned out pretty good. Used a completely different dataset this time. EPS version is stronger than VPred, though it produces a foil effect on hair sometimes.
Pony:
It is what it is. Made super quickly using tags from rule34.xxx. Seems to work.
(FOR VERSION 1.0) Don't use bright VAEs to not get flashbanged for now. Also, the style doesn't really seem strong enough... Might also curate the tags more and remove many duplicate images (there were a lot) :))
Always appreciate seeing your generations too
I still don't know what I'm doing with training settings :3
Description
This time remembered to use a more neutral VAE + save all epochs
Also trained more epochs
The style now resembles the real thing more (i think)
Other things that i changed:
trained more epochs at lower learning rate
now has caption dropout and shuffle caption options enabled
removed trigger word + images that were bad for training
some other little things here and there
For a future version:
will remove 2015 images so it fits the current style more
i dunno... you tell me what to do