hxngiee/DiffFace

model performance

Bruc3Xu opened this issue · 6 comments

Bruc3Xu commented

it seems that the model didn't work well when source and target face have different mask orientation.
output_i_0_b_0

hxngiee commented

As shown in the figure below, it basically works even if the source and target have different poses.

The reason for incorrect samples seems to be attributed to the following factors:

  1. ID loss is set too large
  2. Target-preserving blending time is set too short.

It is recommended to lower the loss weight in the next line or decrease the time duration through masking_threshold argument.

output_i_0_b_0

Bruc3Xu commented

thanks for your reply, I changed the parameter setting including IDloss weight, mask_threshold, but still cannot get good result.The follwing images were used for testing.
004
004_target

luoxyhappy commented

same question, do you solve it now?

akankshasingh25 commented

Were you able to solve it @luoxyhappy @Bruc3Xu ? The face swaps are not of good quality even after manipulating loss parameters and hyperparameters.

luoxyhappy commented

Were you able to solve it @luoxyhappy @Bruc3Xu ? The face swaps are not of good quality even after manipulating loss parameters and hyperparameters.

Not yet.... And i have already given up trying now.

akankshasingh25 commented

Not yet.... And i have already given up trying now.

Same, I have been at it for a few months now. I couldn't find a way yet. @hxngiee, any help would be appreciated.