step1:

With vision model and language model,  the description of the picture are converted into clay style prompt.

step2:

With flux controlnet and lora, we can transform the style of the image while maintaining similarity.

step3:

Crop and enlarge the face, and redraw the face to maintain similarity through sdxl and instant id models.

主要分三步：

1、通过视觉模型和大语言模型，实现prompt 的风格转换

2、通过Flux/lora/controlnet，在转换风格同时保持相似度。

3、通过对脸部采取放大重绘再贴回，这一步使用sdxl模型和instantid模型，来恢复脸部相似度。

flux lora:

https://www.liblib.art/modelinfo/9e0e6f29c0a245d3bb74bad9c7e2313e?from=personal_page&versionUuid=e0087451768a4bb3844bd084ebd79962

sdxl checkpoint:

https://www.liblib.art/modelinfo/36404bfb74f64c168cad048312eb5f42?from=personal_page&versionUuid=3414d7c35b1245a3b3bfcf11c980cd5b

sdxl lora:

https://www.liblib.art/modelinfo/235e784744304cc9833e6eaad5d9b58b?from=personal_page&versionUuid=29269f9228b84c8cad139a15c3372551

手机端H5在线使用链接（flux模型叠加sdxl模型会有点慢）：

https://tt.9syun.com/app/index.php?i=66&c=entry&do=mobile&m=tech_huise&r=index.index.index#/pages/zhixing/zhixing?id=3155&tg=0&v=2