同样是小红车的高热度二游，阴影、打光、肉感和皮肤质感都相当出色，感觉和上一个交错战线的lora契合度不错：暖色肉感肌肤和油丝的搭配。学习率很低，泛用性还是很可以的。数据集不光学了游戏CG还学了部分角色的立绘
色调比较鲜艳，感觉这游戏的角色CG都有点自发光？肉感皮肤、glossy skin
线上训练好贵哦，希望大家多点反馈和图图，喜欢的话来点赞给咱回回血
I really need your like！
触发词：@browndust、glossy skin
trigger word：@browndust 、glossy skin
使用方法：
- 推荐正向前缀：masterpiece, best quality, score_9, score_8, highres, absurdres, @browndust, ...
- 推荐负向：worst quality, low quality, score_1, score_2, score_3, artist name, blurry, jpeg artifacts
- 推荐权重：0.9（可按喜好调 0.6-1.0）
- 采样：30 步、CFG 4、Euler（simple 调度）、1024x1536、ModelSamplingAuraFlow shift 1-3
Usage:
- Suggested positive prefix: masterpiece, best quality, score_9, score_8, highres, absurdres, @browndust, ...
- Suggested negative: worst quality, low quality, score_1, score_2, score_3, artist name, blurry, jpeg artifacts
- Recommended weight: 0.9 (adjust 0.6-1.0 to taste)
- Sampler: 30 steps, CFG 4, Euler (simple scheduler), 1024x1536, ModelSamplingAuraFlow shift 1-3
- Works on any character; style also transfers to non-training characters
Training details:
- Base model: Anima Base v1.0
- Network: LoRA (networks.lora_anima), dim 64 / alpha 64, UNet-only
- Dataset: 215 images x 2 repeats (430 per epoch), resolution 1536x1536, bucket 1024-1536
- Optimizer: AdamW (weight decay 0.01), LR 2e-5, constant scheduler, warmup 0
- Timestep: logit_normal, weighting logit_normal, discrete_flow_shift 3.0
- Loss: L2, gradient checkpointing, bf16, cache latents
- 50 epochs = approx 10,750 steps
Description
- v2：训练到 41 轮，相比 v1 减轻过拟合，同时光影削弱了一点
- v1：存在轻微过拟合，多 LoRA 叠加时容易出现坏脚趾/坏手指。
- v2: trained to 41 epochs (~8,815 steps). Reduced overfitting compared to v1.
- v1: slightly overfitted; when stacked with other LoRAs it could
produce bad toes / bad fingers. This v2 fixes that while keeping the style and
glossy stocking quality.