tip1:注意步数再20到64即可，过低的步数出图会扭曲，过高的步数会出现过拟合，

tip2:然后是lora的权重最好再0.4~0.7之间，太低出图就不像中国人了，太高就很像原图没有随机性，

可以参考我做的图的参数

最后，记得点赞，点赞，点赞呀，

Tip 1: Pay attention to the step count between 20 and 64. A step count that is too low may cause distortion in the generated image, while a step count that is too high may result in overfitting.

Tip 2: The weight for LORa should preferably be between 0.4 and 0.7. If the weight is too low, the generated image may not look like a Chinese person, and if it's too high, the generated image may lack randomness. You can refer to the parameters I used for the generated image.

Finally, please remember to like, like, and like!

补充：

best quality, ultra high res, (photorealistic:1.4), masterpiece, Kodak portra 400, film grain,blurry background ,indoors,1girl, sleeveless turtleneck, (busty), ((huge breasts)), large breasts, (brown long hair:1.3), (looking at viewer), (full body), (closeup) <lora:Chinese doll idol 方脸 square face-000006:0.7>

Negative prompt: (NSFW:2),paintings, glossy skin, wet, sketches, (worst quality:2), (low quality:2), (normal quality:2), lowres, normal quality, ((monochrome)), ((grayscale)), skin spots, acnes, skin blemishes, age spot, glans

Steps: 20, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: 3982801562, Face restoration: CodeFormer, Size: 512x768, Model hash: fc2511737a, Model: chilloutmix_NiPrunedFp32Fix, Clip skip: 2, ENSD: 31337

2.0更新

重新修图，剔除了一些图，修改训练的lora步数，上传2.0。

步数28时，不开高清修复lora权重在0.5~0.68,开了高清修复,lora最大权重可以＋0.1

2.0 Update:

The images have been re-edited and some of them have been removed. The number of Lora training steps has been modified, and version 2.0 has been uploaded.

With 28 steps and no high-definition restoration, the Lora weights ranged from 0.5 to 0.68. With high-definition restoration enabled, the maximum Lora weight can be increased by 0.1.

补充：

对不起，各位，我不知道vae和采样方法的重要性，所以没写，现在补上。、

大模型：chilloutmix

vae:vae-ft-mse-840000-ema-pruned

采样方法：DDIM

Here's the translation:

Sorry everyone, I didn't mention the importance of VAE and sampling methods earlier, so I'm adding them now.

The name of the large model is "chilloutmix", and it uses the VAE model "vae-ft-mse-840000-ema-pruned" with the sampling method "DDIM".