My most powerful model to date, which went through multiple iterations and a lot of re-merges. There's not much else to say about it. I aimed for a 2.75D style of sorts, mostly realistic, but with a capacity (and propensity) for vivid coloration, nice light/shadow interplay and an ability to understand both anime tags and plain text.
The biggest flaw I've encountered so far while using this little beast is that it has a hard time generating sex scenes. Other than that, I really enjoy using it.
Recommended settings are as follows:
Sampler: DPM++ 2M Karras
Sampling steps: 20~30
VAE: the one that comes with SD 1.5
CFG Scale: 7~11
Excessively long negative prompts tend to make the output worse, not better, and most negative embeddings do more harm than good too. Keep the negative prompt nice and simple.
Description
Used a new, more guided approach to merging layer by layer as opposed to the more "blind" method I used to employ that's basically "eyeball the recipe and hope it works".