No trigger word needed.
Recommended settings:
LoRA strength: 0.75–1.0
Step: ≥16
CFG: 3–5
Sampler: Euler/Heun
Scheduler: Simple
The training set is taken from OrangeO recent miscellaneous image collection, with a portion of the bad-hand images removed; resolution 1344 px.
For tagging I used the official recommended JSON format (long, medium and short descriptions included; on top of that I also marked the censor positions with bboxes — maybe that helps prevent the model from overfitting to the censored parts?).
See: https://github.com/OpenMOSS/MOSS-Image/blob/main/USERGUIDE.md
Follow OrangeO, and thanks for the follow.
https://www.pixiv.net/users/128894200
不需要触发词。
推荐设置：
LoRA strength: 0.75–1.0
Step: ≥16
CFG: 3–5
Sampler: Euler/Heun
Scheduler: Simple
训练集取自橘佬近期杂图集，剔除了一部分坏手的图片，分辨率 1344 P。
打标方面我使用了官方推荐的 JSON 格式（包括长中短三种描述，除此之外我还用 bbox 标出了 censor 的位置，也许能防止 censor 部分过拟合？）
详见：https://github.com/OpenMOSS/MOSS-Image/blob/main/USERGUIDE.md
关注橘佬喵，关注橘佬谢谢喵。