
Three versions based off of different training methods. The "orig(ingal)" is straight up screenshots from the show, "full" (forgot why I named it this) was reduced size to 512x512 focusing on Konata, and "redux" was a reduction of images from "full" that had parts of other characters.
I personally find the redux version to be best of the three, although it gets confused about the mole. It is supposed to be under the left eye, but tends to end up under the right. If it isn't there, go into ms-paint, choose a black marker, and add it manually.
This does work with clip skip = 1, but no pictures were added below because I think they looks bad. Weights range from .75-1, play with it as needed. Fairly compatible with other TIs and LoRAs, some issue here and there.
For all those interested, all images were taken after the 18th birthday party (episode 17 first 5-ish minutes). Trained on novelai checkpoint at a rate of .0005, all data listed with the original
描述:
Resized training set to 512x512 images. This version is okay.
训练词语: konatafull-10000
名称: konatafull-10000.pt
大小 (KB): 46
类型: Model
Pickle 扫描结果: Success
Pickle 扫描信息: No Pickle imports
病毒扫描结果: Success