search

megvii-research / CADDM

Official implementation of ID-unaware Deepfake Detection Model
Apache License 2.0
145 stars 21 forks source link

Cross dataset testing #50

Open 13864084796 opened 1 month ago

13864084796 commented 1 month ago

Hello, thank you for sharing the code. I would like to ask, after retraining for 80 epochs, the AUC of the model on the FF++dataset is around 0.98, but the AUC across datasets on the CELEB-DF-V2 dataset is only around 0.74, which is very poor. May I ask why this is the case?

13864084796 commented 1 month ago

I trained on FF++c23, does it have anything to do with that?