这篇文章主要是对GAN的一些论文的总结。
1.Domain Translation with Conditional GANs:from Depth to RGB Face-to-Face
结论:
1.这个是用mseloss , 我们在用loss的时候是不是可以考虑用l1smoloss,或者mse
2.dnet输出是1通道,我们可以选着1通道来做
3.这里mse的倍数是10倍
2.Conditional GANs For Painting Generation
![image.png](https://upload-images.jianshu.io/upload_images/13874392-d8e9f7c04e52f95d.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)
结论:
1.这里在D网络中引入很多信息来增强网络,这是不是表示网络学习能力差,或者鉴别能力差,我们需要引入参数来训练D网络
2.对合并后的信息都需要进行重新编码,这样才不会把特征当成一种
3.网络的学习率是0.00002,这里是不是要搞清楚,对于GAN训练中,我们怎么来调整两个网络的学习率。
3.WAV2PIX: SPEECH-CONDITIONED FACE GENERATION USING GENERATIVE ADVERSARIAL NETWORKS
1.这里在D网络插入了语音信息,这有什么用??
2.结果不怎么样,但是这样的网络也能训练出来,可以看看网络参数,和训练参数。
3.G=0.0001,D=0.0004
4.ICface: Interpretable and Controllable Face Reenactment Using GANs
1.
5.Hierarchical Cross-Modal Talking Face Generation with Dynamic Pixel-Wise Loss
结论:
1.训练参数:0.0002
2.mse的参数是10倍
6.Mask-Guided Portrait Editing with Conditional GANs
7.Reconstructing faces from voices
1.lr=0.0002
8.Cycle In Cycle Generative Adversarial Networks for Keypoint-Guided Image Generation
lr = 0.0002
9.GP-GAN: Gender Preserving GAN for Synthesizing Faces from Landmarks
9.High-Quality Facial Photo-Sketch Synthesis Using Multi-Adversarial Networks
10.Triple consistency loss for pairing distributions in GAN-based face synthesis