Updated: Jul 22, 2026
character
Krea 2 Version
Trained using a fresh dataset built with my workflow. It uses 90 images. 30 realistic. 30 anime. 30 realistic-stripped down. By "stripped down" I mean all aspects of her outfit and accessories were removed. This mean this iteration of this lora has what we'll call modularity. And individual piece of her outfit can be removed, though all of them will also be required in the work flow.
But by far my most versatile lora, and a good case study for how I'll handle them going forward.
Flux Klein 9b Version
These were trained as Rank 64, so they are HEAVY. But they are also quite versatile. I also did not use outfit tags here, since with FK9 I was having trouble putting them in anything else. You can be broad about the outfit and get her normal outfit, in this case "white red and gold halter top, black skirt with purple sash, red leggings with gold design" will get it done.
Works great for T2I and Image Edit.
Ernie Image
All images here are generated on turbo. The model is pretty versatile, using a 43 image dataset and has style flexibility. IF the outfit isn't showing up properly use the tag. I did not do anything with her accessories in the dataset captions, so you might not be able to remove her glasses or headdress.
Qwen 2512 Coming Soon
Z Image Turbo
Trained on 5k steps because I had trouble with certain details, so it can be a bit wonky. It helps to prompt her outfit (top and dress separately)