Practice — Generative Adversarial Networks: From Minimax to StyleGAN (6 questions)

Your team is building a real-time avatar feature: a user's webcam feed is style-transferred into an animated character face at interactive frame rates (tens of milliseconds per frame). A teammate proposes reusing the company's existing diffusion-based image pipeline, arguing "diffusion gives better quality and we already have the infrastructure — let's just use a very small number of denoising steps."

1. Using the five-axis comparison table from this subject (latency, quality, diversity, training stability, controllability), make the case for a GAN-based approach instead.
2. Is the teammate's "just use fewer steps" proposal a real fix for the latency problem? Explain why or why not, tying your answer to what DDIM-style step reduction actually changes.
3. Name one axis where the teammate's diffusion-based proposal would still have a real advantage over a GAN here, and explain why that advantage doesn't outweigh the latency requirement in this case.

You're training a face-generation GAN from scratch. Early in training, the discriminator's loss drops sharply and stays very low, while the generator's loss stays high and essentially flat — generated images remain obviously blurry and unrealistic for far longer than expected, with almost no visible improvement epoch over epoch.

1. Explain the minimax objective's two competing terms in plain language, and state which one the generator is trying to minimize.
2. Given the symptom described, what specific problem with the textbook minimax generator loss is the likely cause, and why?
3. What is the standard fix, and why does it address the specific mechanism from part 2 rather than just being "a different loss that happens to work better"?