- Eliminate classifier, discriminator acts as classifier
- Use discriminator only for real/fake detection
- Without Autoencoder Training
- With Autoencoder Training, to ensure encoder captures all information required for reconstruction
- With Variational Autoencoder Training
- Separate Decoder, Generator
- Decoder == Generator
- Approach 1a: Random Noise, Expected Fake
- Approach 1b: Inverted Image, Expected Fake
- Approach 2: Random Image belonging to class c2, Expected c2
- Approach 3a: Random Noise, Expected Mismatch
- Approach 3b: Inverted Image, Expected Mismatch
- Approach 4: Inverted Image, Expected Inverted