Images in showcase are base gens, no hirez, no facefix. Of course it looks much better fixed see gallery for fixed images.

If DPM++2M burns ure image (ex. close ups) switch to Euler A

CFG 5->15 maybe less maybe more!

When refining this model from dpm karass gen -> Euler A I use 5%-10% for whole pic, 15-25% for face)

Euler A - with CFG values 10-15 is totally viable way to run this model

This model is kinda hard to work with, but its a sdxl merge experiment so it is what it is, however:

THIS MODEL IS A VERY GOOD REALISTIC REFINER MODEL FOR PONY OR OTHER MODELS. Just refine with dpm++2m karass on like 40-70% denoise and you will get very good results.

So this is a merge between Raver which is a very slight refine of base pony, and the araminta experiment cv6 - cause I like that model, especially with my crazy lora filter that I have used on it. The model is semi realistic but not a lot of that was added tbh, a lot of other sources got somewhat nuked in the merging process though but with added loras it can do styles quite well. If you want other source the realistic I recommend the Raver model instead, its more versatile.

For duo ???? you gotta prompt wisely. Use extra weights if needed. Use guider loras. It can do it alone but failure rate is high.

It has trouble with eyes a bit. glowing eyes in neg can help

It has a lot of trouble with hands - "long nails" in neg. And use many samples and/or add loras to help with refinement.

Nipples can easily become a problem "inverted ???????" in neg. Another way is to add loras that handles breasts in someway.

Will add all prompts later if I feel inclined.

The pink haired girl in black armor, has a refiner lora stack on that seemed pretty good. You will not need all of these loras. Just a few if you want base output to be more stable is enough. The more loras at low values though the more refinement without impacting overall style too much