Version 9: should work with most samplers. I dent to use CFG 4-5 and 30 steps. Sometimes the pony triggers can be helpful like score_9, ...DMD2 versions: quite fun and super fast LCM Karras or Normal CFG 1.2-1.5, 4-8 steps You can use slightly higher CFG like (8 steps or more). Hires. fix I usually use 4 steps.Version 4: Works with DPM++ SDE starting 6 steps but 9 is better (6 is a bit soft) and cfg 3 - 3.5 Works with other sampler too but did not test those much yet. They seem ok and work at fairly low steps too.V3 is a complete rework, all samplers are usable. I am not sure what is going on model hash is the same but I seem not to be able to recreate the same pictures. I don't know if the windows update or Nvidia update messed something up for me.V2 was made to be used with DPM++ 2M SDE (karras or exponential), some others should work too. 3M SDE and 2S a seem to work well too. Steps 12-16 works fairly well. You can push CFG fairly high to like 12 but might depends on steps.As an experiment merges some lightning and others and 3 of my Loras.Quite like it so I thought I would share.DPM++ SDE Karras starting 6 steps and CFG 2 works (not sure if there is limit 20 steps cfg 5 works too for example).Model is a bit horny so I guess I picked the right name ;)

Centerfold

Partially merged a new Lora I trained on top of 3.5

centerfold_v38.safetensors

v3.8

version XI dmd2

version X (dmd2)

v9.0 naughty

v8.0 dmd2 naughty

v7.1 dmd2

v7.0 dmd2

v6.0 dmd2

v5.0

v4.0 flash

v3.8 pcm dpo

v3.5

v3.0

v2.0 t5

v2.0

v1.1

v1.0

Files with SHA256: 58c3cdd802d81568df497a07d37aed2e9a7de1225331416dba5cd0481c6213da