Generative AI Models Explained, From GANs to Diffusion

• Generative AI reached 53% population adoption within three years, faster than the PC or the internet, per the Stanford HAI 2026 AI Index.
• US private AI investment across all of AI totalled $285.9 billion in 2025, per the same April 2026 report.
• Score-based SDE models reached FID 2.20 on CIFAR-10 in November 2020, per Song et al.
• Carlini et al. extracted over a thousand training images from diffusion models in their January 2023 USENIX Security paper.
• Discrete diffusion language models run up to 10x faster than autoregressive decoding in the best cases reported by a 2025 survey.

What Generative AI Models Are and How They Differ from Discriminative Models

A generative model is a statistical model that learns the distribution of its training data well enough to produce new samples from it. That’s the whole idea. Chatbots, image generators, and molecule designers are all built on top of it.

The Stanford CS236 course notes state the goal directly: approximate the data distribution, given access to a dataset. Treat every training example as a draw from some unknown process, then fit a model that behaves like that process. The standards-body version says the same thing in different words.

Generative AI is “the class of AI models that emulate the structure and characteristics of input data” in order to generate derived synthetic content. (NIST Computer Security Resource Center glossary, sourced from NIST SP 800-218A)

Two capabilities follow from learning a distribution.

• Sampling: draw a novel datapoint from the model. This is what people mean when they say a model “generates”.
• Density estimation: ask the model how probable a given datapoint is under what it learned. Handy for anomaly detection and for checking whether an input resembles the training data at all.

One terminology point before going further. “Generative model” is the statistical concept. “Generative AI” is the application-level term for products built on such models. Marketing copy uses them interchangeably, and that blurs a line worth keeping.

Generative versus discriminative is the contrast that actually explains the name. A discriminative model such as logistic regression learns p(y|x): given an input, predict the label. A generative model learns the joint distribution p(x,y) over inputs and labels together. From the joint, Bayes’ rule gives you p(y|x), so a generative model can classify. The reverse derivation is impossible.

The last row comes from Ng and Jordan’s 2001 NIPS comparison of naive Bayes and logistic regression. The discriminative model wins on asymptotic error. The generative model reaches its own, higher, error level with far less data. A February 2023 arXiv paper revisiting the pair summarises the result the same way.

Here’s the caveat most explainers skip. That finding covers one simple model pair from 2001. It is a poor basis for a blanket rule like “generative beats discriminative on small data”. A 2016 paper on implicit modeling adds nuance: generative models tend to resist overfitting, while discriminative models are more powerful and less exposed to misspecification once data is plentiful.

How a Generative Model Learns to Produce New Data

A generative model learns by choosing a family of candidate distributions, then adjusting that family’s parameters until the model’s distribution matches the training data as closely as the family allows. The Stanford CS236 introduction to deep generative models describes the parametric case exactly that way: learning means picking the parameters within a model family that best match the data.

In practice the “family” is a neural network. Its weights define a distribution. Change the weights and you change which outputs the model considers likely.

Every generative model family shares the same loop: pick a parametric family, fit its weights to the data, then sample from it or score inputs with it. Source: Stanford CS236 course notes.

The fitting step is ordinary gradient-based optimization against an objective that scores mismatch. Each family defines that objective differently, and those differences are what separate a VAE from a GAN from a diffusion model. The loop itself is shared.

Randomness has to enter somewhere. A trained model is deterministic, so to produce something new it needs a source of variation on the input side. That source is usually a simple distribution the model transforms into something complex: Gaussian noise, a latent vector, or a random pick among plausible next tokens. OpenAI’s undated post on generative models frames the payoff of this setup as forcing the network to discover the data’s most salient features, because it can’t reproduce the distribution without them.

Sampling and density evaluation are different operations:

1. Sampling runs the model forward from noise to output.
2. Density estimation holds an input fixed and asks the model to score it.
3. Some families make one cheap and the other expensive. Many make both tractable, but at a cost in sample quality or speed.

Anyone who has fine-tuned a pretrained model has run this loop in miniature. You keep the family, start from someone else’s parameters, and nudge them toward a narrower distribution. That’s why LLM fine-tuning is cheap relative to pretraining: the hard part of matching the broad distribution is already done.

Types of Generative AI Models: Transformers, Diffusion, GANs, VAEs, and Flow Models

The main types of generative AI models are autoregressive models (including Transformer-based LLMs), variational autoencoders, generative adversarial networks, normalizing flows, diffusion or score-based models, and flow matching. Each is a different answer to the same question: how do you represent a complicated distribution so that a network can learn it and sample from it?

Autoregressive models treat a sequence as a chain of conditional probabilities: predict the next token given everything before it. GPT-3, described in the May 2020 paper “Language Models are Few-Shot Learners”, scaled this to 175 billion parameters and was evaluated few-shot with no gradient updates. Every large language model in production today descends from this recipe.

VAEs, from Kingma and Welling’s December 2013 paper, introduce a latent variable and train on a reparameterized lower bound that stochastic gradient methods can optimize. GANs, from Goodfellow et al. in June 2014, skip likelihood entirely. A generator tries to fool a discriminator, and in theory the game ends with the generator recovering the data distribution and the discriminator outputting 1/2 everywhere. In practice you almost never see that clean equilibrium, which is the gap between the paper and the training run that frustrates everyone who tries it.

Normalizing flows (Rezende and Mohamed, ICML 2015) keep every transformation invertible, so exact density evaluation comes for free. Diffusion models, introduced by Sohl-Dickstein et al. in March 2015, went quiet for five years and then took over image generation.

Five months after DDPM reported FID 3.17 on unconditional CIFAR-10, Song et al.’s score-based SDE framework reached 2.20 on the same benchmark. Source: Song et al., 2020.

Two 2020 results mark that takeover. Ho et al.’s Denoising Diffusion Probabilistic Models reported an FID of 3.17 on unconditional CIFAR-10 in June 2020. Song et al.’s score-based SDE framework from November 2020 unified score matching with diffusion, hit FID 2.20 on the same benchmark, and recovered exact likelihood through a neural ODE. Flow matching, from Lipman et al. in October 2022, then generalised the approach: it trains continuous normalizing flows by regressing vector fields along fixed probability paths, treats diffusion paths as a special case, and reported better likelihood and sample quality than diffusion baselines on ImageNet.

One point gets lost in all the architecture names. The category is defined by what gets modeled. Whether the network is a Transformer, a U-Net, or a stack of invertible layers is an implementation choice, and the same family can sit behind text, images, audio, or molecular structures.

What Are Generative AI Models Used For in Practice?

Generative AI models are used to produce new text, images, video, audio, and molecular structures, and to score how likely a given input is under the data they were trained on. The first use gets the headlines. The second is quieter and often more useful inside a company.

The NIST glossary definition lists images, video, audio and text as the “derived synthetic content” a generative model produces. Jun Lu’s book-length survey of generative models, published on arXiv on August 8, 2026, adds molecular structures and describes models that now cover all five. Same model families, different data.

What that looks like as actual work:

• Drafting and completion: autoregressive language models write one token at a time. GPT-3’s May 2020 paper reports that human evaluators had trouble telling its generated news articles from human-written ones.
• Image, video and audio synthesis: diffusion and flow-based models run from noise to a finished picture, clip, or waveform.
• Few-shot task adaptation: GPT-3 was evaluated with examples in the prompt and no gradient updates, so one pretrained model can be pointed at a new task without a retraining run.
• Screening and scoring: density estimation lets a model flag an input that looks nothing like its training data, which needs no generation at all.
• Molecular design: sampling candidate structures from a model trained on known molecules, as the 2026 Lu survey documents.

What I see in enterprise projects is less glamorous than that list. Most production work is text, and most of that is summarising, extracting, or drafting against documents the company already owns. The image and video modalities draw the demos. The text modality draws the budget.

Adoption has outrun every prior computing wave. The Stanford HAI 2026 AI Index, published April 13, 2026, reports that generative AI reached 53% population adoption within three years, faster than the PC or the internet. The same report puts total US private AI investment in 2025 at $285.9 billion.

That $285.9 billion covers all of AI. The Index summary gives no separate generative-AI-only line, so any gen-AI-specific investment figure you see quoted elsewhere deserves a second look before it goes in a slide.

Generative AI reached 53% population adoption within three years, while US private investment across all of AI totalled $285.9 billion in 2025. Source: Stanford HAI AI Index, 2026.

How to Choose the Right Type of Generative AI Model for a Task

Choose a generative AI model by matching five facts about the task to the strengths of each family: output modality, whether you need exact probabilities or just good samples, the latency budget, how much training data exists, and how much control you need over the output. Modality narrows the field fastest. The other four decide between the survivors.

Latency is the criterion teams underweight. A diffusion sampler that looks fine in a notebook becomes the slowest hop in a request path once it sits behind an API, because each sample runs the full chain of denoising steps. A GAN pays once. If the output has to come back inside a user’s attention span, that difference decides the architecture before anything else does.

Exact likelihood is the criterion teams overweight. Most product features only need good samples. You need a tractable density when you want to compare two inputs, detect outliers, or audit what the model thinks is probable, and in those cases flows or the score-based SDE route are the honest choices.

Resist the “generative models work with little data” shortcut. The Ng and Jordan result covers one simple model pair from 2001. For a deep model, the practical answer to scarce data is to start from a pretrained checkpoint in the right family and spend the budget on evaluation instead of training from scratch.

Why Generative Models Hallucinate, Collapse, or Memorize Training Data

Generative models hallucinate, collapse, or memorize because each one is an approximation of the data distribution, and every approximation fails in a way that mirrors how it was trained. Three failure modes cover most of what goes wrong in production.

Hallucination in autoregressive models. An LLM samples each next token from a learned conditional distribution. That procedure scores plausibility under the training data. Truth never enters the calculation. Fluent output is the design goal, so the failures are fluent too. The GPT-3 paper from May 2020 shows both sides at once: evaluators had difficulty separating its news articles from human-written ones, while the model still struggled on some datasets and the authors flagged concerns about training on large web corpora. A model that writes convincing news will write convincing errors.

Mode collapse in GANs. The minimax game from Goodfellow et al.’s June 2014 paper guarantees that the generator recovers the full data distribution only at equilibrium, where the discriminator outputs 1/2 everywhere. Training rarely gets there. Short of equilibrium, the generator can keep the discriminator fooled by producing one narrow slice of the data well, and it has no incentive to cover the rest. Samples look sharp and varied in a batch of ten. Across a thousand, the repetition shows.

Memorization in diffusion models is the counterintuitive one, because the family with the best image quality is also the leakiest.

Carlini et al. extracted over a thousand training images from diffusion models with a generate-and-filter pipeline, and found them far less private than earlier generative models such as GANs. (Extracting Training Data from Diffusion Models, arXiv, 30 January 2023; USENIX Security 2023)

If your training set holds patient scans or customer documents, that finding changes the deployment conversation. Sampling is no longer a one-way operation.

For a structured checklist, NIST’s AI 600-1 Generative AI Profile, approved in July 2024 as a voluntary companion to the AI Risk Management Framework, identifies 12 risks unique to or made worse by generative AI. It is cross-sectoral and reads well as an audit template.

Mitigations, briefly:

1. Run extraction-style tests for memorization before a model touches sensitive data, instead of after.
2. Ground text generation in retrieved documents, which is the practical reason RAG pipelines exist.
3. Track sample diversity across large batches for GANs, since a small batch hides collapse.
4. Keep a human check on any output that will be read as fact.

What Does Training or Running a Generative Model Actually Cost?

Training a generative model costs compute in proportion to parameter count and data volume, and running one costs compute on every request in proportion to how many sequential steps a single output needs. Those are two separate bills. Teams budget the first one carefully and then get surprised by the second.

Training cost scales with size. GPT-3, in the May 2020 paper, had 175 billion parameters, and that number is still the reference point people reach for when they want to say “large”. Every parameter gets touched on every training step. Pretraining at that scale is a project for a handful of organisations, which is why nearly everyone else starts from a pretrained checkpoint and pays only for adaptation.

Inference cost is where the architecture matters.

Sequential decoding is the autoregressive tax. An LLM cannot write token 200 until token 199 exists, so long outputs mean long waits and a GPU held for the whole time. Discrete diffusion language models produce tokens in parallel, and a 2025 survey of discrete diffusion in language and multimodal models reports quality comparable to autoregressive models with inference up to 10 times faster. Read that 10x as the strongest case the survey authors found. Head-to-head numbers on one fixed benchmark are thin.

Image diffusion has the same problem in a different shape. Each sample pays for every denoising step, so step count is the cost knob, and turning it down trades quality for speed.

Fine-tune or call an API? Prices change too fast for a dollar figure to survive publication, so think in shapes. An API converts cost into per-token spend with zero training outlay and no control over the weights. Fine-tuning front-loads a smaller training bill and hands you a model you host and pay for per request. The crossover depends on volume. High, steady traffic favours owning the model. Spiky or exploratory traffic favours renting it.

How Generative Models Changed in 2026

Generative models changed in 2026 in three ways: diffusion began competing with autoregressive decoding for text, flow matching settled in as a unifying training method, and the field started treating text, image, video, audio and molecules as one subject. Adoption numbers moved as well.

Diffusion for language is the shift with the most practical weight. Autoregressive decoding has owned text since the 2017 Transformer paper. Discrete diffusion language models denoise an entire sequence in parallel instead, and the 2025 survey of discrete diffusion in large language and multimodal models reports results on par with autoregressive models at up to 10x faster inference in its best cases. A second survey devoted to diffusion language models carried a June 2026 version, so the literature is moving month to month. Whether parallel denoising displaces next-token prediction in production remains open. It has at least become a real option.

Flow matching grew from a 2022 paper into a default recipe. Lipman et al. showed in October 2022 that regressing vector fields along fixed probability paths covers diffusion as a special case, with better likelihood and sample quality than diffusion baselines on ImageNet. By 2026 that generality is the selling point: one training method, many paths, and no need to commit to a noise schedule up front.

Multimodal generation became the standard framing. Jun Lu’s book-length survey, published on arXiv on 8 August 2026, treats models that produce text, images, video, audio and molecular structures under a single set of principles.

Adoption figures from the April 2026 AI Index:

• Generative AI reached 53% population adoption within three years of arrival.
• Singapore leads at 61%.
• The United States sits at 28.3%, ranked 24th.

The United States sits at 28.3% generative AI adoption, ranked 24th, well behind Singapore’s leading 61%. Source: Stanford HAI AI Index, 2026.

The US figure surprised me. The country that trains most of these models ranks in the twenties on using them, and that gap between building and adopting is the number worth watching into 2027.

Frequently Asked Questions About Generative Models

These are the questions people actually search about generative models, each answered in a few sentences that stand on their own.

Is ChatGPT a generative model?

Yes. The chat product is an application, and the model underneath it belongs to the autoregressive family: it learned a distribution over text and samples from it one token at a time. The May 2020 GPT-3 paper describes that recipe, a 175-billion-parameter Transformer that adapts to new tasks from examples in the prompt with no gradient updates.

Is a generative model the same as generative AI?

No. A generative model is the statistical object: a learned approximation of a data distribution that you can sample from or evaluate probabilities under. Generative AI is the product-level term for systems built on those models, which is the sense NIST’s glossary uses when it describes AI models that emulate input data to produce synthetic content.

Are large language models generative or discriminative?

Generative. An LLM learns the joint distribution over token sequences by factoring it into next-token conditionals, and sampling from that chain is what produces text. A discriminative model only learns to map an input to a label and could never write a sentence.

Which generative model is best for images?

Diffusion, score-based and flow-matching models hold the strongest published results. Song et al.’s score-based SDE model reached FID 2.20 on CIFAR-10 in November 2020, and Lipman et al.’s flow matching beat diffusion baselines on ImageNet in 2022. GANs remain the fastest at sampling because they generate in a single pass, so the answer depends on whether quality or latency matters more for the job.

Do generative models copy their training data?

Sometimes, and diffusion models do it most. Carlini et al. showed in January 2023 that a generate-and-filter pipeline could pull more than a thousand training images back out of diffusion models, which they found far less private than GANs. Test for extraction before deploying on sensitive data.

Can a generative model be used for classification?

Yes. A model that learns the joint distribution p(x,y) can compute p(y|x) through Bayes’ rule, which is exactly how naive Bayes classifies. Ng and Jordan’s 2001 comparison found such a classifier settles at a higher error floor than logistic regression while reaching it with fewer examples, a result that holds for that model pair and should be stretched no further.

Where to Start Building with Generative AI Models

Start building with generative AI models by picking one modality and one task, loading a pretrained model from the family that fits, and testing it for memorization and hallucination before anyone depends on it. Training from scratch comes later, if ever.

The order matters:

1. Name the modality and the job. “Text, summarise claims documents” narrows the field to autoregressive models in an afternoon. “Something with AI” never narrows it at all.
2. Begin from a pretrained checkpoint. The broad distribution is already learned, so your budget goes to adaptation and evaluation.
3. Run extraction and factuality tests early. Carlini et al. pulled more than a thousand training images out of diffusion models in 2023, and the cheapest moment to discover that a model leaks is before it holds customer data.
4. Read the originating papers. Kingma and Welling (2013), Goodfellow et al. (2014), Rezende and Mohamed’s ICML 2015 paper on normalizing flows, Sohl-Dickstein et al. (2015), and Lipman et al. (2022) are short, and they explain the failure modes better than any vendor deck.

A generative model is an approximation of a distribution. Hold it to that standard, measure where the approximation breaks, and most of the drama drains out of the project. The first working version should be small, boring, and measured.

If you’d rather skip the false starts, AlphaCorp AI’s generative AI development team builds and tests these models against your production data from the first week.