Pillar 1 · Model Selection — family deep dive

Types of GANs

Every GAN is the same competition — a generator forges samples, a discriminator calls out fakes, both improve until forgeries pass. The types differ in what conditions the generator: nothing, a class label, another image, a low-resolution input, or a table schema. That conditioning is what you're actually buying.

The competition, once

Generator vs discriminator

You never need the game theory (the "Nash equilibrium") — you need to know the generator is a one-pass, deployable artifact, and that training GANs is famously unstable (an ML-team problem, but it explains why off-the-shelf beats train-your-own).

The types, by what conditions the generator

Type	Conditioned on	Named models	Choose it when
Unconditional GAN	Random noise only	DCGAN (the classic), StyleGAN2/3	Endless realistic variations of one domain — faces, textures, avatars ("this-person-does-not-exist")
Conditional GAN (cGAN)	A class label	BigGAN, conditional StyleGAN	You need to say what to generate — "a defect of type B" for dataset balancing
Paired image-to-image	An aligned input image	pix2pix	Sketch→photo, map→satellite — when you have matched before/after training pairs
Unpaired translation	An image from another domain	CycleGAN	Day↔night, summer↔winter, style domains — no matched pairs needed (great for augmenting camera data)
Super-resolution	A low-resolution image	SRGAN, ESRGAN, Real-ESRGAN	Upscaling archives, video enhancement, zoom-and-enhance pipelines — the most-bought GAN today
Tabular / synthetic data	A table schema + real rows	CTGAN, TVAE (SDV toolkit)	Privacy-preserving datasets for analytics, testing, sharing — synthetic customers instead of real PII
Stability variants	— (a training fix, not a product)	WGAN, WGAN-GP	Trivia worth recognizing: "Wasserstein loss" fixed vanishing gradients — the canonical example of math you can skip and the MLE cannot

Where GANs still win vs diffusion: single-pass speed (real-time upscaling, low-latency generation), mature niche pipelines (ESRGAN in video tooling), and non-image data synthesis (CTGAN for tables). For text-to-image quality and control, diffusion has taken the crown.

Governance note: the same generator that balances your defect dataset can fabricate convincing fakes. Synthetic-data programs need provenance tracking (label generated data as generated) — auditors and the EU AI Act will ask.