Type | |
Stats | 156 |
Reviews | (33) |
Published | Nov 11, 2023 |
Base Model | |
Training | Epochs: 7 |
Usage Tips | Clip Skip: 2 |
Trigger Words | Taishi hearts sparkles blush fang blush stickers |
Hash | AutoV2 4CBA321283 |
Much higher character fidelity! This version is--i dare say--100 times better.
Added noise offset
Eliminated examples in data set causing pollution
Still flexible on outfits and backgrounds, but can actually make Taishi's actual outfit now.
Should play nicely with other LoRa's. I weighted it at 1 for the examples, but it can be toned down. It still makes the character and only loses the outfit a bit at around 0.5-0.7. Wont matter if you want to make him wear something else anyway.
Post any suggestions or generations you make!
Enjoy!
_____________________________________________________________________________________________________
Attempted to make a LORA for Taishi from Drifting Home and thought I would share what I have made in case anyone else wants it. I am still fairly new to this, so I am making changes to these models as I learn better methodology. Of course, I welcome any feedback people wish to provide as well!
Model is of a character I don't own, so probably don't do any of the commercial use stuff.
I can't say what weight will work best due to very different results per base model. Sometime, weighting the trigger word and turning down the LORA weight works better. I intend to update this once I figure out how to train the model to understand Taishi's outfit better. For now, simply using "jacket" should get close on some base models, but the colors can be awry.
use "Hearts" and "sparkles" tags for some more pizzazz and, while I havent tested it, "chibi" should have some effect based off the data set. The model is pretty receptive to changes like hair color, eye color, freckles, or you can even make Taishi into a pretty little girl by adding long hair to your prompts...though he might have something to say about that.
Hmm...I doubt anyone is still reading or cares too much, but none of the sample images are upscaled. I Just saved some decent looking generations to use here from a few different models.
I am considering doing these next:
Ashe -- Fire Emblem Three Houses
[Holding a Nintendo Switch] *the AI doesn't understand this well enough to my liking
Cain Von Silford -- Chronicles of an Aristocrat Reborn in Another World
Trombone LORA ???? IDK if this will work, but I am going to try