In general I used the same (manually decensored) source material as for my other (PONY) LoRAs. This time around, I trained it locally.
Version 2 is completely re-tagged and uses new training params. Also character names should (mostly) work.
Version 3 was completely retrained and pruned (cutting of dimensions that bled context). This should result in fewer visual errors, more adherence to the prompt and clearer texts.
You can extract the ComfyUI workflow from the example pictures. Just download them and drag those into your ComfyUI webview.
Base keyword is "v1p3r1se".
Additional character keywords are:
Karin
Makoto Takanami
Akira Honjouji
Lilia Milcrave
Saki
Narsha
Irear
Raika Grace (technically, not RISE series, but fit the style quite well)
Description
I completely changed the training of this version and pruned several dimensions to minimize the LoRA checkpoint, as well as minimizing the concept bleeding. Details should now be better, less mangled. Texts should be more readable.
FAQ
Comments (3)
Really nice
💕👍🏻
Wow, you used my favorite character persona (IAyita: the short hair girl with glasses and hoop earrings).... nice