search

cuiaiyu / dressing-in-order

(ICCV'21) Official code of "Dressing in Order: Recurrent Person Image Generation for Pose Transfer, Virtual Try-on and Outfit Editing" by Aiyu Cui, Daniel McKee and Svetlana Lazebnik
https://cuiaiyu.github.io/dressing-in-order
Other
513 stars 127 forks source link

Face is different from the original image #32

Closed kieutrongthien closed 2 years ago

kieutrongthien commented 2 years ago

How i can increase quality this image ?. Result is too different from original image. image

kieutrongthien commented 2 years ago

How i can increase quality this image ?. Result is too different from original image. I'm using DIORv1_64, this is my custom image.

image

cuiaiyu commented 2 years ago

The released model was trained on DeepFashion, and unfortunately, there is some overfitting on faces (as DeepFashion only a few dozens of different faces). You may refer to those Virtual Try-On methods which combines the garment and the other part of the person on image level, in which the face is copied directly from the source images.

kieutrongthien commented 2 years ago

The released model was trained on DeepFashion, and unfortunately, there is some overfitting on faces (as DeepFashion only a few dozens of different faces). You may refer to those Virtual Try-On methods which combines the garment and the other part of the person on image level, in which the face is copied directly from the source images.

So, your mean is if i prepare the dataset with more different faces and results will be improved right?

cuiaiyu commented 2 years ago

Yes, a more diverse set of faces would help the collapse/overfitting situation.