The Goal of this Checkpoint was to get a similar base style like my previous models. This could not be achieved by mixing some Artists and even with the Prompt Scheduling node I was not able to come even close.So I decided to train my own one - after a lot of research I ended up by training 4 different Loras, which are based on my own generations and other artstyles I liked. For training data I used ~1 000 hand picked Images (~250 for each style, mostly with women and men + some background images --&gt; especially with my choice of men I did something wrong, the all had rather big members...).The Captioning was made with a custom Qwen 3 VL model and a custom prompt within comfyui. The captioning was in pure natural language and only described the artstyle not the content of the image - so no trigger word needed. Somehow the results where good at first sight and after some deep testing I decided to make some other approaches - this was very time and energy consuming...Then I mixed the loras into the basemodel with different weights - so technically it is a merged model - with lots of own training. I'll release some of the loras later on with thier own distinctive style.The new checkpoint comes much closer to the base style of my old models. I really like to play with different artstyles which work suprisingly well - you can use my <a rel="ugc" href="https://civarchive.com/models/2018003/artists4semireal">Artist4Semireal</a> wildcard to try yourself. Also other loras work.Feel free to experiment and post some generated images here - feedback is always welcome. You can add the Turbo- and Highres-lora from Circlestone Labs - which work well together, if you tame down the Highres one.

Files with SHA256: f65b030404c760f7506702bebeccc5afa5153441b0af263ecfa1723b0063e1eb