search

10Ring / LAA-Net

The official implementation for LAA-Net: Localized Artifact Attention Network for Quality-Agnostic and Generalizable Deepfake Detection
Other
55 stars 4 forks source link

Train on FF++ C23 #4

Closed wuwuwuyue closed 2 months ago

wuwuwuyue commented 2 months ago

We tried to train on FF++C23, but the training results were very poor. The AUC on CDFv2 was less than 70%. Here is the config file attached. It may be a problem with our processing. Has anyone successfully reproduced the results? 屏幕截图 2024-07-12 204425 屏幕截图 2024-07-12 204353

sbi.json

10Ring commented 2 months ago

Hi @wuwuwuyue , thanks for your interest in our work.

Our results are reported by training on the c0-raw version of FF++. We did not train on c23. In my opinion, working with c23 is a good try; however, besides try consulting our preprocessing steps, please also try to experiment some other things, e.g. hyper-params when you do SBI, new data augmentations, training (lr, etc). Many thanks.