The Apothecary Diaries (Kusuriya no Hitorigoto 薬屋のひとりごと) || Maomao / Jinshi / Gyokuyo / Lihua / Xiaolan / ...
详情
下载文件
关于此版本
模型描述
所有模型均为捆绑格式,仅能在 WebUI 版本 >= 1.7.0 上使用
参见 https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13568
中间检查点:https://huggingface.co/alea31415/random_stuff/tree/main/kusuriya
本模型使用我更新的流水线 https://civitai.com/articles/2383,该流水线自动从动漫中构建数据集。
通配符:
针对版本 1(第 01-03 集):https://huggingface.co/alea31415/random_stuff/blob/main/kusuriya/EP01-03/kusuriya.txt
针对版本 2 和版本 3:
Jinshi
Jinshi, long sleeves, bun cover, wide sleeves
Maomao
Maomao, long sleeves
Maomao double bun
Maomao double bun, earrings, hair ornament, necklace, hanfu, long sleeves, makeup
Maomao ponytail
Maomao ponytail, long sleeves, japanese clothes
Maomao cat ears
Maomao cat ears, long sleeves, hair ornament, blush, wide sleeves
Lishu
Lishu, earrings, necklace, hair flower, hair ornament
Gyokuyou
Gyokuyou, earrings, hair flower, necklace, hair ornament
Ah-Duo
Ah-Duo, earrings, kimono
Lihua
Lihua, earrings, hair flower, cleavage, hanfu, makeup
Yinghua
Yinghua, earrings, hair flower, hair ornament
Guiyuan
Guiyuan, hair ornament, hair flower
Fengming
Fengming, hair ornament, kimono, japanese clothes
Maomao_hair_down
Maomao_hair_down, earrings, sparkle, lipstick
Fuyou
Fuyou, kimono, earrings, hair flower
Gaoshun
Gaoshun, hat, long sleeves, wide sleeves
Ailan
Ailan, earrings, hair flower, hair ornament, hanfu
Hongniang
Hongniang, earrings, necklace, hair ornament
Guen
Guen, bun cover
Lingli
Lingli, blush, baby, hair flower, ^_^, hair ornament
ServantA
ServantA, long sleeves
Luomen
Luomen, japanese clothes
TasterA
TasterA, earrings, hair ornament, hanfu
Xiaolan
Xiaolan, hair bow, hair rings, long sleeves
Lihaku
Lihaku, hat, sweatdrop, long sleeves
MadamK
MadamK, hair ornament, earrings
Pairin
Pairin, hair flower, cleavage, large breasts, fur trim
screenshots
booru
事实证明,部分角色的训练效果不佳。我特别建议对 Yinghua 和 Ailan(可能还包括 Guiyuan 和 Hongniang)使用版本 1。否则,添加 screenshots 可提升细节保真度,但会牺牲灵活性。(实际上,对于某些角色,过多内容会被它吸收。)
与版本 1 不同,雀斑现在能正确关联到 Maomao。即使在双髻造型下,雀斑也可能出现。此时在负面提示中加入“无雀斑”可能有所帮助。
服装的灵活性始终是个问题,尤其是对于图像数量较少的角色。
我将 OFT 设为默认检查点,并非因为其必然优于 LoHa,而是为了展示这种同样有效的方法。我使用了更低的学习率(OFT 为 2e-5,LoHa 为 2e-4)。最终来看,OFT 检查点似乎略具更强的灵活性(例如,使用 PVC 模型时,LoHa 可能需要配合 LBW,而直接使用 OFT 则无需调整),但这不应被视为两种方法之间的有意义比较。




















