|
Model | Mechanisms | Advantages | Disadvantages | Applicable scenarios |
|
GAN | Generator and discriminator | Generate high-resolution images | Training instability, mode collapse, vanishing gradient | Generating partial images |
CGAN | Add additional information to the input layer of the generator and discriminator | Effectively constrain the overly free GAN | Requires labeled data, training imbalance, the quality of generated images is low | Generating specified target images |
DCGAN | Combine GAN with CNN, use BN and other techniques to train the model | Rich variety of generated images | The quality of generated images is low and model training is unstable | Generating most images |
WGAN | Use Wasserstein distance instead of JS divergence in traditional GAN | Prevents GAN training instability and mode collapse | Unreasonable parameter settings can easily lead to gradient dispersion | The GAN model does not converge and the mode collapses |
LSGAN | Use the least squares loss function instead of a traditional cross entropy loss function | Generate high-quality samples | The gradient vanishes or explodes during training | Generating high-quality images |
BigGAN | Expand the scale of the model, use truncation and orthogonal regularization to train | Model training is stable and can generate highly clear images | Large number of parameters and difficult to train | Suitable for generating highly clear images |
|