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
517 stars 127 forks source link

Bad results by the pre-trained model when tested on an unseen image #62

Closed onchiptech closed 1 year ago

onchiptech commented 2 years ago

Hello,

I run the "Try-On" task inference with the pre-trained model on the following image which is not in the dataset.

a

As you can see the below output generated is not good. Face and hands are not rendered properly.

Screenshot from 2022-09-23 12-56-06

But for the images in the datasets, I am getting good results.

Any suggestions?

Regards, Surendra

cuiaiyu commented 2 years ago

The pretrained model has the face overfitting issue for outlier data. For try-on task, I'd suggest something like directly copying and pasting the face as

final_img = face_mask * input_img + (1-face_mask) * generated_image.

where the face_mask you can get from the human parsing which is part of the inputs.

vinodbukya6 commented 1 year ago

Hi @onchiptech

Did you solve the face overfitting issue?

When I used this method, Results are not good. 

final_img = face_mask * input_img + (1-face_mask) * generated_image.