Misato Katsuragi LoRA for SDXL
Trained with: Kohaku XL beta

Summary:
This model is quite good at adding Misato to your gens, while also not destroying the original knowledge of the checkpoint. This should also make it remarkably flexible given exotic prompts, however the strength may need to be increased beyond 1.

Info:
This LoRA is the result of some experimentation that led to some very interesting insights. I have been putting together new iterations of my Misato Dataset, and have been testing various settings for d_coef and weight decay when training with Prodigy. I ended up with 4 decent-quality LoRAs that I was comparing, when I realized that it might be possible to merge these all into the same LoRA. I used the LoRA merge tool in Kohya's GUI to mix those 4 LoRAs at various weight strengths into a single LoRA. Interestingly after messing around with tuning the weights, the new LoRA did a much better job of maintaining a consistent character while preserving the original generation subjects of the model (mostly).

This model is capable of SFW and NSFW.

• The "original" tag should increase the appearance of a 1990s anime screencap, and this dataset includes only SFW official art.

• A more fanart style is triggered by "misatofanart", and this dataset includes SFW+NSFW fanart.

• Please experiment with this network, post your results, and feel free to let me know if you have any issues or suggestions!