Hanfu Song/宋制汉服SDXL

SDXL宋制汉服LORA，女生版。

使用指南：

1、触发词为szhf dress。不过这个触发词可有可无，使用后权重会更高一些。

2、建议权重设为0.6-0.8；

3、支持宋制汉服的各种特殊样式。（不添加以下细分词则随机生成。）

在关键词中添加“beizi”可生成褙子；“short beizi”生成短褙子；“wide sleeve”生成大袖；“narrow sleeve”生成窄袖；“shan”生成交领衫或裙掩衣的上衣；“belts”生成腰带；“shawl”生成披帛，“xiapi”生成霞披；“pleated skirt”生成百褶裙。

此外，可通过添加“with a guan on her head”生成宋冠（但宋冠出图概率较低，因数据集中冠的图片较少，仅顺带训练了一下）。

4、我初步测试后感觉泛化性尚可，大家可以参考我发布的测试图，均为模型直出、未经高清修复或放大的结果。虽然这是写实风格模型，但与其他模型配合，也能生成二次元或插画风格的图像。

5、横图、竖图等各种比例均支持，但因主要使用3:4比例训练，3:4比例可能效果最佳；

6、本模型主要针对汉服，面部和手部表现一般，生成全身图时建议开启After Detailer。（后续我可能会对这部分进行微调。）

7、最后，欢迎同学们多分享返图和提出建议，比心！

——————————————————————

How to use it:

1. The trigger word is szhf dress. You can skip it, but using it gives the style more weight.

2. I recommend 0.6-0.8 for the weight.

3. It supports kinds of Song dynasty hanfu styles. (It'll randomly generate stuff if you don't specify.)

You can add “beizi” in the keywords to generate 褙子; “short beizi” for 短褙子; “wide sleeve” for 大袖, “narrow sleeve” for 窄袖; “shan” for 交领衫或裙掩衣的上衣 ; “belts” for 腰带; “shawl” for 披帛, “xiapi” for 霞披; “pleated skirt” for 百褶裙.

Additionally, you can add "with a guan on her head" for 宋冠， (but the probability of generating 宋冠 is lower since there are few images in the dataset. I just lightly trained it on them for fun).

4. From my initial testing, the generalization seems OK. Please refer to the test images I've posted，I didn't use upscale.

Although this is a photorealistic model, using it with other models can generate anime or illustration style images.

5. It can generate various aspect ratio images, but 3:4 may work best since that's primarily what it was trained on.

6. This model focuses mainly on Hanfu clothes. Faces and hands may not be ideal, so I recommend enabling After Detailer for full body images for now. (Or I may fine tune these parts later.)

7. Finally, I welcome everyone to provide more test images and feedback! Appreciate it! love you all!