Worst-Case Morphs Using Wasserstein ALI and Improved MIPGAN

Figure 4

Comparison of morphs generated by a 128  128 WALI model trained with FR losses (top row), a 128  128 WALI model trained without FR losses (second row) and a 512  512 WALI model without FR losses (bottom row). Comparing the top two rows shows that there seems to be a trade-off between image quality and morphing performance. Comparing the two bottom rows shows that blurriness can be corrected by simply generating higher-resolution images.