Figure 4. Challenge videos of original and deepfake samples. Each row aligns outputs against the same instance of challenge, while each column aligns the same deepfake method. This illustration also shows the driving videos of imposter.
Figure 5. Artifacts defined for human evaluation. (a) has no deepfake artifacts, (b) has a boundary artifact near the right brow, in (c) the hand vanishes behind the face, (d) has hazy face detail, and (e) the sunglass starts vanishing briefly. Red objects indicate artifact locations..
Figure 12. Sample challenge images obtained using a adaptive variant of DFL trained with diverse data for long training time along with a segmentation module. As observed this version is able to do several challenges with lesser artifacts.