johannwyh / HifiFace

This is the official project website of RealFace: 3D Shape and Semantic Prior Guided High Fidelity Face Swapping
59 stars 4 forks source link

Questions of SFF #2

Open linyu0219 opened 3 years ago

linyu0219 commented 3 years ago

First,nice work! Results are unbeliveable. I have two questions of SFF module. 1.In figure 3(b) SFF details,there is an Adain Resblock between Zfuse and Fup ,Zfuse must be an input of the Adain ,what's the another input?

2.In figure10 ,arctile say

To further illustrate how SFF addresses the problem, we show difference feature maps of every stage in the SFF module, named SFF-0∼3

What's the output in figure 3 are SFF-0∼3?

Thanks for your answering.

Last, Great Work Again.

johannwyh commented 3 years ago
  1. The another input is same as that in Decoder, the 3D Shape-Aware Identity . You can refer to Figure 12 in the supplementary material of our arxiv version for detailed structure.
linyu0219 commented 3 years ago
  1. The another input is same as that in Decoder, the 3D Shape-Aware Identity . You can refer to Figure 12 in the supplementary material of our arxiv version for detailed structure. Thanks for your answer. But I can't understand which is SFF 0-3 in figure 12 ,either.
johannwyh commented 3 years ago

Sorry, forget to comment on question 2. In our analysis experiment,SFF-0~4 corresponds to two AdaIn after feature map blend and two res blocks in F_up.

linyu0219 commented 3 years ago

Sorry, forget to comment on question 2. In our analysis experiment,SFF-0~4 corresponds to two AdaIn after feature map blend and two res blocks in F_up. I see,Thanks for responding