search

ldz666666 / RiDDLE

Author implementation of RiDDLE: Reversible and Diversified De-identification with Latent Encryptor (CVPR 2023)
43 stars 2 forks source link

Recover for real image #4

Closed Harxis closed 1 year ago

Harxis commented 1 year ago

Thank you for your excellent work. I noticed that you actually use the w-code generated images to do the comparison of the recovered effect during the training and testing, why don't you use the recovered image to do the comparison with the original real image?

ldz666666 commented 1 year ago

During training, we calculate the loss between the inverted image (w code generated) and the generated ones, because they are at the same domain. You can paste the generated face back to the original background using a pretrained face parsing network in celebamask-hq.