Introduction to Generative AI – Use Cases and Types

TL;DR

Generative AI learns the underlying pattern (distribution) of training data well enough to produce new, original samples that look like they came from the same distribution — new images, text, audio, code, whatever the domain is.

Why it matters

Almost every "AI tool" headline since 2022 — ChatGPT, Midjourney, voice clones, code copilots — is a generative model under the hood. Understanding what family a tool belongs to tells you what it's good at and where it'll break (a GAN is great at sharp images but unstable to train; a diffusion model is stable but slow at inference; an autoregressive model is great at sequences but generates one token at a time). This topic is the map — each family gets its own deep-dive later in this course.

The core idea

Simple shabdo mein: a discriminative model answers "what is this?" (classify), a generative model answers "give me something new like this." It does that by learning (explicitly or implicitly) the probability distribution P(x) of the training data — or the joint P(x,y) if labels are involved — and then sampling from that learned distribution to create new data points that were never in the training set but are statistically plausible members of it.

The hard part is that real-world data distributions (natural images, fluent language, music) are extremely high-dimensional and complex. Different model families take different mathematical routes to approximate and sample from that distribution.

How it works

At a map level, the major families this course covers are:

• VAE (Variational Autoencoder) — compresses input into a probabilistic latent space z, then decodes z back into data. Learns an explicit, tractable lower bound on P(x) (the ELBO). Fast to sample, but outputs tend to be a little blurry since it optimizes a bound, not the true likelihood.
• GAN (Generative Adversarial Network) — two networks, a Generator and a Discriminator, play a minimax game against each other until the Generator's fakes are indistinguishable from real data. Produces sharp samples, fast single-pass generation, but training is notoriously unstable (mode collapse, vanishing gradients — covered in later topics).
• Autoregressive models — generate data one piece at a time, each piece conditioned on everything generated so far: P(x) = Π_i P(x_i | x_1,...,x_{i-1}). GPT-style language models are the classic example. Excellent quality, but generation is inherently sequential and slow.
• Diffusion models — start from pure noise and iteratively denoise it over many steps until a clean sample emerges (the reverse of a fixed noising process). Extremely stable to train, very high quality (DALL-E 2/3, Stable Diffusion), but slow at inference since it needs many denoising steps.
• Flow-based models — learn an exactly invertible transformation between a simple distribution (e.g. Gaussian) and the data distribution, so the exact likelihood P(x) can be computed (no bound needed, unlike a VAE). Mathematically elegant, exact density estimation, but the invertibility constraint limits network architecture choices.

Each of these — except flow-based models, which stay at this map level in this course — gets its own dedicated topic later with full architecture and maths.

The maths

No new formula at this map level — see the canonical sheet in later topics (VAE's ELBO, GAN's minimax value function, diffusion's forward/reverse process) for the full derivations.

Worked example

Quick gut-check you can use in the exam: if asked "which family would you use to generate a 10-second clean denoised image fast," the answer is GAN (single forward pass through G). If asked "which family gives the most stable training at the cost of slow sampling," it's diffusion. If asked "which family gives an exact, tractable likelihood," it's flow-based (VAE only gives a lower bound on likelihood).

Use cases across domains

One concrete example and one real challenge, per domain:

• Healthcare — generating synthetic medical images (e.g. MRI/CT augmentation) to train diagnostic models when real labeled scans are scarce. Challenge: a hallucinated anatomical detail in a synthetic scan could mislead a model trained on it, so clinical validation is essential.
• Education — generating personalized practice questions and explanations adapted to a student's level. Challenge: factual accuracy — a generative model can produce a confident but wrong explanation (hallucination), which is actively harmful in a learning context.
• Entertainment — AI-generated game textures, NPC dialogue, or background music. Challenge: maintaining stylistic consistency across a whole game/film without manual art-direction oversight.
• Marketing — generating ad copy and product images variations at scale for A/B testing. Challenge: brand voice consistency and avoiding generic/repetitive output across campaigns.
• Software development — code-completion and code-generation copilots (autoregressive language models). Challenge: generated code can be subtly incorrect or insecure while looking syntactically fine, so human review stays necessary.
• Content creation — text-to-image tools for illustration/concept art. Challenge: copyright and provenance of training data, plus the output occasionally containing visual artifacts (extra fingers, warped text) that need a human pass.

Exam traps

• Don't call GANs and VAEs "the same kind of autoencoder" — a plain autoencoder isn't generative in a useful sense (no structured latent sampling), a VAE is.
• Autoregressive ≠ diffusion. Both generate step-by-step, but autoregressive conditions on previously generated output tokens, diffusion conditions on progressively less noisy versions of the whole sample.
• Don't say diffusion models are "fast" — they're known for quality/stability, not speed, because of the many denoising steps.
• Flow-based models are the only family here with an exact tractable likelihood — VAEs only get a lower bound (ELBO).

Remember this

Generative AI = learn the data distribution, then sample new points from it — the model families differ only in how they represent and sample that distribution.