EDIT:
Version 3 is here. I added 100 new sci-fi themed images and re-trained model with the entire dataset (800 images), 30000 steps, 8 epochs.
CFG sweetspot - 4
Version 2:
This is a pure Finetune, trained in 2 iterations so far, 12 epochs and about 20000 steps on nvidia RTX A6000. I will definitely continue to work on it. I have (intentionally) generated all images without high-res fix, so that you can have a better insight into the basic capabilities of this model.
Negative prompts are not necessary, i advise you to experiment with them.
In addition to captions, the model is also trained with tags. which have strong impact on image generation. For example, a large part of the dataset consists of images of ninja and samurai warriors, and they are tagged with: ninja style, samurai style, ninjutsu, katana, buddhist.
1st iteration: dark style, darkness, fantasy, gothic, horror
2nd iteration:
fantasy style, dark, surreal, horror, epic
ninja style, samurai style, ninjutsu, katana, buddhist
CFG sweetspot: 4-6
U can support my work here:
https://www.buymeacoffee.com/nenadkuzmao