Suggested settings:
Positive: <lora:betty_gelpin_lora_v1:1> betty_gilpin, 1girl, solo, blonde hair, long hair, looking at viewer,
Negative: (empty to start, add nsfw if desired)
Clip Skip: 1
Resolution: 768x512
pretty tolerant of other settings - 7 cfg, 20 steps, 0.2 denoise for high res if you use it.
CLIP SKIP 1 is MUCH BETTER on some models, and slightly better on most models. If you really need clip skip 2 for some other reason, I'd suggest running it through a second pass at clip skip 1 just doing the face with the suggested starting prompt.
164 images, 28k steps, mostly 768x512, but many 512x512 and 512x768 as well.
Trigger word: betty_gilpin
Most common tags in data set: 1girl, solo, blonde hair, long hair, looking at viewer,
I tried a couple times to get this right, honed in on just the right amount of training at a very low learning rate for 3.5 hours on a 3090 ti. Removed the word 'realistic' from the tags in the training set. Few tags needed for realism and few negative tags needed as well. I suggest you start with just the trigger word and popular tags and start from there. Add nsfw into the negative prompt if you want to ensure that.
If you find you need them (if you're getting deformations and crazy stuff), here are a few negative embeddings I liked (one at a time, of course):
ng_deepnegative_v1_75t
BadNegAnatomyV1-neg
negative_hand-neg
bad_prompt_version2-neg