Four generative model types with real-world applications

What's actually being asked

Pick any four generative model families, explain each briefly, and give one concrete real-world application per model.

Answer

The four picked below cover genuinely different generative mechanisms — latent-variable, adversarial, iterative-denoising, and autoregressive — rather than variants of the same idea, which is what this kind of question is usually checking for.

1. Variational Autoencoder (VAE)

An encoder maps input x to a distribution q(z|x) = N(μ(x), σ(x)²) over a latent space; a decoder reconstructs x from a sampled z. Trained by maximizing the ELBO (reconstruction accuracy + a KL term keeping the latent space close to a standard normal prior). Application: drug discovery / molecule generation — a VAE trained on known molecular structures can sample new, chemically plausible candidate molecules from its latent space for virtual screening.

2. Generative Adversarial Network (GAN)

Two networks are trained in a minimax game: a generator G tries to produce fake samples realistic enough to fool a discriminator D, while D tries to tell real from fake — min_G max_D 𝔼[log D(x)] + 𝔼[log(1 − D(G(z)))]. Application: image super-resolution — upscaling low-resolution photos or satellite images to high-resolution, sharp outputs (e.g., in photo-editing tools and remote sensing).

3. Diffusion model (DDPM)

Learns to reverse a fixed forward process that gradually adds Gaussian noise to data; a network ε_θ(x_t, t) is trained to predict the noise at each step, and generation runs this reverse process from pure noise back to a clean sample. Application: text-to-image generation — tools like Stable Diffusion generate high-quality, diverse images from a text prompt, now widely used in digital art, marketing creative, and product design mockups.

4. Autoregressive Transformer (large language models)

Models the joint distribution of a sequence as a product of conditionals, generating one token at a time: P(x_1,...,x_n) = Π_t P(x_t | x_1,...,x_{t-1}), using self-attention (Attention(Q,K,V) = softmax(QK^T/√d_k)V) to condition on all previous tokens. Application: conversational AI assistants and code-completion tools (e.g., chatbots, AI pair-programming assistants) — generating coherent text or code one token at a time conditioned on context.

Why these four, together

VAEs and GANs are both primarily used for continuous data (images) but differ in training stability vs. output sharpness; diffusion models now dominate high-quality image generation precisely because they avoid GANs' adversarial instability; autoregressive transformers dominate discrete sequence data (text, code) where "generate one step at a time" is the natural framing. Seeing all four side by side is usually enough to answer most "compare generative model types" questions on this paper.

Exam tip

When asked for "any four," pick models from genuinely different mechanisms (not two GAN variants, say) — it signals breadth. Pairing each with a crisp, specific real-world use (not just "used in image generation") is usually where marks are lost or won.