This LoRA aims to produce a fantasy style similar to the one used in the Diablo Series. Most of the training data was concept art from the games, some 3D art (mostly of nude Lilith because that's what you find if you look for it) and a few in-game screen captures I did from Diablo IV.
The checkpoint I used for the example images is a custom merge I did with several LoRAs and my KestenXL. No, I currently don't plan on releasing it. I simply do not see the point. Checkpoints on this site are very oddly picked for the onsite image generator and I am yet to figure the logic behind. So far I'm thinking it is: quality=irrelevant, moderator/admin gluteus maximus tongue polishing=maybe.
Also, this site is packed full of checkpoints. I'm pretty sure you don't need my low-effort, simple and boring merges. Just choose some of the abundant hand-picked ones available for on-site generation and let your imagination run wild.
Since Diablo (especially the 4th installment) is intimately related and connected with The Mother of All, Lilith, I thought she deserved a tag of her own, so:
1) "lilith" - 2D art of Mama L;
2) "lilith-3d" - 3D art of Mama L.
The "wings" tag produces demon, bat-like wings (un?)fortunately every time. Reason why? Mama L's doting, smothering presence in the dataset, and my, low-qualitylingly and large-sizeably forgetting to invent a tag like "light wings" or "wings of light", or "angel wings" for the angels' wings made of light. Darn, I guess. Typical.
UPDATE (IL version):
Now we have a proper tag(s) for wings - "light wings" to get the angelic wings, and "demon wings" to get the bat-like demon ones.
See example images to know what to expect from either of those two. Other than that, everything else, including Mama L's two tags is in t he dataset. +1 like if you want me to make a StarCraft and or WarCraft LoRAs like this one.
Description
This is a second training attempt, because I was unhappy with how the first one came out. So saddened by its deficiencies was I, that I chose not only to not release it, but also to retrain it twice - once as a LoRA and once as a LoCON.
This is the LoRA (better dataset and with a trigger word). I will not release the LoCON (it's better in my opinion) because it has the stunningly gargantuan size of 308 MB, and of what I've seen, people on this site are very sensitive to size. At the end of the day, SIZE doEs iNdEeD maTTer. Who would've thunk...