Hello everyone!
Important!
I recreated a version 1 "universal" of the model that should be compatible with many more models. I had trained the base model in "clip skip -3" but this brings too many problems with models other than mine. I of course leave the other version available for those who want to test it with my models.
This new LoRa model was designed to give a more realistic side to my Katayama_Mix_XL v1 and v2 models, but also works very well with SDXL_Niji_v6 and a little less well with Seven even if we can do things by not putting too much Strength on it
It was created with a mix of images from MidJourney v6 and the video game Hellblade 2.
So the eyes are more detailed, the skin textures are more present, the level of detail is increased and the contrasts are better.
The results are optimal in "clip skip -2" (clip skip -3 for SDXL_Niji_Seven and Katayama_Mix_v2") and with the sampler "dpmpp_sde"
The sweet spot of the model is around 0.8 but in some cases with Katayama_Mix can work at 0.9 or even 1.0.
Have fun! 😊