Test of test of test
trigger words for v10: miwasirou
trigger words for v24 and Anima v1.0: 5h1r0wm1wa
A lora based on the art style of artist Shirow Miwa.
This lora was actually trained a year ago, initially just for testing, using only 10 sampled images. The results were surprisingly good. Subsequent training with many more images failed to match this version.
Recently, I trained another version using 24 sampled images. The training results have their own strengths and weaknesses compared to the initial version.
The 10-image version is named "v10," and the 24-image version is named "v24."
v24 is not necessarily better than v10
v10 is very stable and accurately reflects the artist's style from a specific period. However, its stability leads to somewhat stiff character designs and composition, more prompts need to be added to improve the generated results.
v24 is less stable but has a wider range of applications. Character designs and composition are more expressive, and it performs better than v10 in generating complex structures.
Note that the two LoRa have different trigger words.
Anima v1: The learning rate might be a bit low. It works best with a strength of around 1.1 to 1.3. I'll update it later if necessary, but I think it's basically usable now.
Why are descriptions getting longer and longer......
