1. Disadvantage of GANs

Advantages

• Amazing empirical results
• Fast inference

Disadvantages

• Lack of intrinsic evaluation metrics
• Unstable training → mode collapse
• No density estimation → unable to get the likelihood of a particular image
• Inverting is not straightforward

2. Alternatives to GANs

2-1. Variational Autoencoders(VAEs)

Variational Autoencoders

1. Real image input to encoder

2. Encoder outputs mean and standard deviation

3. Use reconstruction loss between the fake output of the decoder and the original real input to the encoder

4. Sample from distribution with the outputed mean and standard deviation

5. Take sampled value (vector/latent) as the input to the decoder

6. Get fake sample

1. Backpropagate through

Maximizing the Evidence Lower Bound(ELBO)

• Maximizing the likelihood of real images
• Learned probability distribution thinks that a real image occurs
• likelihood is mathematically intractable but the lower bound is tractable

→ we maximize the lower bound making the likelihood better

$p(z)$ : prior latent space distribution

→ represents the likelihood of a given latent point in latent space