realgirlmix2是基于赛博Dream | CNrealistic_MIX_V40融合和加练得来的，侵权可删

在实践中对比赛博Dreamv40的本体，改进了v40本体融合的lora脸型，改善了半身以上的光影，使得光影更加具有层次感，增加了出图的暖色调，并改善了面部脸型。

注：但是手部直出等稍微比赛博dream本体好一点因为素材使用了大量的带有手部的人物图像。

在线生成麻烦用实际tag不要用嵌入式。

可以再本地使用(worst quality:2),(low quality:2),(normal quality:2),lowres,watermark,badhandv4,ng_deepnegative_v1_75t,

一定要开高清修复，一定！！！

支持768x768的出图能力

下面推荐参数是我复制原模型的，通用

推荐参数：

•vae：已经嵌入vae-ft-mse-840000-ema-pruned 、

•采样：DMP++2M Karras（强烈推荐） ，1.6新采样器也可尝试

•步数：30~50

•高分辨率修复：20步以上

•放大算法：4x-UltraSharp、8x_NMKD-Superscale_150000_G等搭配使用。

•重绘幅度：0.2~0.5

•放大倍数：1.5倍以上。

Realgirlmix2 is based on CyberDream | CNrealstic_ MIX_ Derived from V40 integration and refinement, infringement can be deleted

In practice, by comparing the ontology of CyberDreamv40, the Lora face shape integrated with v40 ontology has been improved, and the light and shadow above the body have been improved, making the light and shadow more layered, adding warm tones to the image, and improving the facial shape.

Note: However, compared to the main body of the blog, it is slightly better to have the hands straight out because the material uses a large number of character images with hands.

Please use actual tags instead of embedded ones for online generation.

Can be used locally (best quality: 2), (low quality: 2), (normal quality: 2), lowres, watermark, badhandv4, ng_ Deepnegative_ V1_ 75t,

Make sure to open high-definition repair, make sure!!!

Supports 768x768 graphics capability

The following recommended parameters are for me to replicate the original model, universal

Recommended parameters:

• Vae: already embedded in vae-ft-mse-840000-ema run

Sampling: DMP++2M Karras (strongly recommended), 1.6 new samplers can also be tried

Steps: 30-50

High resolution repair: more than 20 steps

Amplification algorithm: 4x UltraSharp, 8x_ NMKD Superscale_ 150000_ Used in combination with G.

Redraw amplitude: 0.2~0.5

Magnification: 1.5 times or more.