discriminator requires temporal frames of 32 but data contains 16 frames

[Shen-Chenhui]

As discussed in the paper, data are of either 16 x 128 x 128 and 16 x 64 x 64, and 5 discriminator layers with multipliers of (2,4,4,4,4) for the former. However, 5 downsampling discriminator layers require at least 32 frames, does the author use padding or actually disable 1 of the temporal downsampling in the discriminator layer?