复现weshop人台模特换装生成的pipeline
结论: 基于sam+inpainting+lora可以复现weshop的pipeline
weshop的数据要求: 绿幕人台模特
直接给一张带背景网图测试,首先基于SAM分割出所有物体,手动选择衣服的部分,获取mask,然后进行生成:
origin | weshop 1 | weshop 2 |
---|---|---|
![]() |
![]() |
![]() |
得到几个信息:
- weshop只进行了衣服外部区域的重绘
- 重绘时使用了controlnet的edge类约束(测试发现是为了约束衣服的边缘不向外扩张)
- 人脸效果差,应该没有引入lora模型
- 除了生成模型外,没做其他后期处理
Using sd-webui-segment-anything to get cloths mask:
origin | mask |
---|---|
![]() |
![]() |
A bit poor result, may due to low resolution or webui plugin defects.
origin | inpainting 1 | inpainting 2 |
---|---|---|
![]() |
![]() |
![]() |
- different model like sd1.5-inpaint or lora
- with/without controlnet
-
single face swap:
- InSwapper
- ReliableSwap
- HifiFace private model
-
face inpainting like miaoya_pipeline
after cloth inpainting | face swap | face inpainting |
---|---|---|
![]() |
![]() |
![]() |
- lora model from civitai
- origin photo from Taobao
- pipeline extract from sd-webui based on sd-api