About dataset generation

[Mark-Dou]

The original paper said blending data samples are composited using real images in FF++,but each video we extracted lots of frames, how do we select the sourceImages?each video select one image or something?I would appreciate it if you could solve my problems?Thank U

[neverUseThisName]

Just extract all faces and not swap faces from the same person.

[Mark-Dou]

Thanks for replying me, but a new problem appears that frames in the same videos always select frames in another same videos to blend cause the fixed threshold in find_one_neighbor.How do you solve this problem?