My first attempt to train a fictional Character. I used some old generated (sdxl, flux, z-image) images to train seperate LoRAs, then generated several images with Krea 2 by mixing the loras. Used the results and trained again. This was a very long way - the V3 got deleted because of TOS violation - "no real persons", V3 had no real persons involved in training anyway. So I changed the dataset in V4 and double checked with several AI face research tools that no real person has the same face. This was a lot of work and the reason why I want to publish it so that all the hard work was worth it. V4 was trained wit ai-toolkit and automagic v3 + sigmoid as optimizer - the LoRA stics to the realistic style. Maybe I'll retrain with classic setup and AdamW - the last LoRAs adapted other Styles much better. It plays well with other LoRAs - just look at the example images.

Files with SHA256: 9e199ab8719e74e31bd5b091d7133cb56117d1ee710209b159996718926d7c95