This is my personal-use style LoRA that I created so I wouldn't need an 8-style-LoRA stack to get the style I liked. With the release of more advanced models like Anima and Krea 2 it's even more important, as it's the only way I'm able to transfer that style over.
My aim with this is to address the generic-looking oversaturated, flat, glossy, and grainy look of the anime checkpoint models. It does that by blending a number of styles and aesthetics together, giving it a distinct painterly and semi-realistic feel.
Anima version notes:
- Adding "glossy skin" to the negative prompt will give natural skin texture.
- er_sde normal or beta57, 4 cfg, and 30-50 steps works best in my experience.
- LoRA trigger is @bvga_c0re, but it's optional and I prefer not using it.
Krea 2 version notes:
- No LoRA trigger tag is needed.
- Works great at 0.7 strength when mixing with other style LoRAs. Krea 2 seems especially good at this sort of style mixing.
- The appearance of the character was learned along-side the style, and as such it has a strong influence on character anatomy, facial features, and eyes.
- Style is strongest when prompting for female subjects. Occasionally has issues with giving male characters female features.
Illustrious version notes:
My intention was to have a single LoRA that both strongly captured that style while being cleaner and more flexible to use than the 8-lora-stack (like what Krea 2 is now able to accomplish). But due to limitations of Illustrious I've split it into 2 separate LoRAs.
Comparison Chart:
v2.0 is a typically trained style LoRA, which has a softer style and doesn't alter prompt understanding, ideal for flexible use and mixing with other style LoRAs.
(v2.0 is what most people will want to use)
v1.0 is an intentionally "overtrained" model that aims to capture that style as strongly as possible, at the expense of prompt sensitivity. As a result, with v1.0 you'll need to experiment with prompt weights, e.g. (character name:1.3).
(v1.0 was also an experimental proof of concept, so the quality of the dataset wasn't perfect.)
Description
I used what I learned over the 5 months since v1.0 to create more detailed images for use in the v2.0 dataset. This has allowed for better quality, but with a difference in texturing that the detailing introduced.
The LoRA is also gentler in its training (664 vs 2,400 samples), which allows for better prompting and less interference with the base model and other LoRAs.
It's still rank 96 (sorry for the large file size). I tried lower ranks, but this gave the best quality.