Conditional GAN
AC-GAN
Test- to - Image
Conditional Image Synthesis with Auxiliary Classifier GANs
- Main porpose
Generate images employ lable conditioning.
- Loss Func
G: Maximum LC-lS
D: Maximum LS
-
Net Structure
-
Metrics
- Discriminability:Inception Score with trained inception net
- Diversity:MS-SSIM(Used in image compression algorithms)
- Datasets
- ImageNet
Training data :images with label,Unsupervised Learning
Other achitecture of D
Stack GAN
- Main Porpose
Generate high-quality text-to-images in 2 stage GAN - Loss Func
-
Stage 1:
G0: Minimum LG0 (DKL :Minimum the KL devergence between perturbations distribution and the stardeard Gaussian distribution)
D0: Maximum LD0 -
Stage 2
G: Minimum LG
D: Maximum LD
- Structure
- Metrics
- Quantitative:Inception score
- Qualitative:Human evaluation
- Datasets
- Oxford-102
- MS COCO
- CUB
Images with descriptions
- Experiments
- Contrast experiment: v.s. GAN-INT-CLS GAWWN
- Component analysis: - Stage 2 ,-text twice