Updated: Sep 6, 2024
style
Pony
Trained on the same dataset and parameters as Animagine
------------------------------------------------------------------------------------------------------
v2
I classified the artist style into three categories: shs, cpc, msm,
Each of shs, cpc, msm, includes elements of 1girl and solo, so excluding 1girl and solo from the prompt might strengthen the effect.
While tags like "xxxStyle" might be more user-friendly, they did not work well, so we used this method, which is also employed in Pony. The tags were chosen based on the conditions of 1token and 3characters, and they do not carry any other meaning.
Optimizer: Adafactor
LR Scheduler: cosine with restarts
Unet: 7.5e-6 , TE: 7.5e-6
To further subdivide the style of the same artist, it seems that around 30 images are needed for each style.
------------------------------------------------------------------------------------------------------
v1
After fine tuning AnimagineV3.1, I extracted LORA from the model.
I don't think there are any quality issues with Dim16, but I will upload the model before extraction as well.
This LORA is still experimental, and there's room for improvement in the parameters and dataset, but for now, I'll consider it complete.
Recommended schedule type : SGM Uniform