This Lora uses what I found was very close to Karma, Wai-Real-CN v14, for training so I recommend using that model. The quality is actually quite decent now as she has been using a much higher quality camera. So it works fine at even 0.8-1 weight for high likeness. Use less than that for more flexibility.The main keyword is Karma while the other common names are different outfits (sometimes hairstyles). To see how each look, check the example images as I included 1-2 for each name.Use day token when you don't want images indoors because they're mostly yellow lighting. F4c33 for close up face focus, t0rs0 for upper body, f0lb0d only has 3 and they're odd angles. Bl00r1 is just some close ups that are blurry, used to cover face angles. H44nd is a couple images that has hand to face, putting this along with "hand on own face" "hand on own cheek" in negative can prevent them especially while using "Lisa" which both has hand to face, and Cindy which has 1.The reason all the examples has karma in the negatives is that this better helps generate the different outfits and hairstyles because this lowers the priority of all trained images except the token name in the positive prompt. It is fine to use Karma in positive when you want a blend of them all though. But if you try to generate a specific outfit, it's more useful in the negative prompt.And even though the model I'm using isn't the base pony model, I notice that the likeness is much higher using score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, . Without the scores token the faces tend to droop or be oddly wide.

Karma_Luey

This one uses PONY Relustion2 for training which looks great but the model is big so it uses a lot of memory. It also has less quality images than newer ones now.

Files with SHA256: b6cd8782d0ab2f37a248e14b98feadded1939659ca9ae02fe36b7f04389dc1da