训练集：

• 40张照片

• 使用最新的金字塔噪声和梯度累加步数

食用参数：

• 分辨率：512x768 （高清修复x2,重绘幅度0.5）

• 触发词：realistic, 1girl,solo

• 模型：majicmix/chilloutmix（更多模型泛用程度没测，底模用的chilloutmix，majicmix表现更佳）

• 采样器：DPM++ 2M Karras（其余默认）

• 权重：0.7~2（0.7会开始变得不太像）

注意！墙裂建议开启高清修复！特别是在full body的情况下！！！

这个模型只是在我测试新参数金字塔噪声和梯度累加步数的时候炼出来的玩意，loss值在看着没有收敛的情况下结果出乎我的意料（挺满意O.o

我认为非常完美的一个Lora，不同权重的表现几乎完美！

QQ：2085690574

图斯：Aniki的个人资料 | 吐司 Tusi.Art

Training set:

• 40 photos

• Using the latest pyramid noise and gradient accumulation steps

Parameters:

• Resolution: 512x768 (HD restoration x2, redraw amplitude 0.5)

• Trigger words: realistic, 1girl, solo

• Model: majicmix/chilloutmix (other models have not been tested for general use, chilloutmix was used as the base model, majicmix performed better)

• Sampler: DPM++ 2M Karras (others are default)

• Weight: 0.7~2 (0.7 will start to look less realistic)

Note: It is strongly recommended to turn on HD restoration, especially for full body images!

This model was created while I was testing new parameters for pyramid noise and gradient accumulation steps. Despite the loss value not converging, the results were unexpectedly satisfactory (quite satisfied O.o).

I think it's perfect for a Lora, with different weights performing almost perfectly!