V1.1 80 images ,4 batch size ,3 repeat times,remove all complex backgrounds in images ,network_dim 16 -> rounder face

V1 40 images,4 batch size ,3 repeat times,network_dim 32

K-DollFace represents an experimental LoRA training endeavor on the Flux base (FP8 KiJai) model. Rather than aiming for the precise replication of individual figures, this model explores the extraction and fusion of common aesthetic features from multiple character images, resulting in the generation of new faces with a unified style. My core objective is to create a distinct and highly consistent "Doll-like" aesthetic with "Plump and Hydrated Skin"

训练思路：目标是特征融合而非强调某张图片的特征，要避免过拟合。每张图片学习次数降低到 3 repeats，epochs 稍微降低；标注中强化 Doll 风格的标签，以便让训练侧重于这些标签的风格。