LazyEye-XTC Beta 1

The goal for this merge was to have a stable and clean model for lora training with good knowledge about personalities

and also being able to produce natural real-life like output for generation.

It should work with common basic settings like Euler-A, CFG:5, 25 Steps.

For DMD2 use LCM:1 and 12 Steps. If you use (character) Loras with DMD2 lower the character Lora weight to about 0.65.

If you want to use DMD2 with Euler-A, lower the DMD2 Lora weight a bit. Please note that DMD2 is designed for LCM scheduler for best quality but LCM mostly ignores negative prompts. Solution is to use Euler-A or other uniform schedulers with lower DMD2 weight. Writing prompts more precisely may help too (i.e. shirt with no text). Also note that LCM replaces the CFG setting, CFG should be 1 and higher/lower guidance is done via DMD2 lora weight. (up to 1.2).

Its not the strongest in NSFW but it does the basic stuff.

To improve further i need your feedback. Please test and report.

Thank you.

Additional notes:

For example images i simply used the positive prompt of the most popular SDXL gens on this site from this year.

Just the raw model without Loras, Embeddings or negative prompt. Yes i am a lazy fart.

Todo / Upcoming:

Better promptable face expressions and figure poses