Generative Model
Expanded definition
A generative model learns patterns in observed data and uses those patterns to produce new samples. Depending on the architecture, it may model a probability distribution explicitly, learn a latent representation, generate one token or pixel at a time, or reverse a noising process. The main families used by practitioners are autoregressive models, variational autoencoders (VAEs), generative adversarial networks (GANs), and diffusion models. Large language models are usually autoregressive: they predict the next token conditioned on prior context. Diffusion models are widely used for image generation. VAEs are useful when a structured latent space matters, while GANs remain important for understanding adversarial training and image synthesis. A generative model is not the same as a discriminative model. A discriminative model is optimized to predict a label or output from an input, such as whether an email is spam. A generative model can produce a new sample, such as a draft email, an image, or a synthetic training record.
Generative model taxonomy
The families differ in how they represent and sample from learned data patterns. The right choice depends on the output modality, controllability requirements, latency budget, and evaluation method.
| Family | How it works | Examples | Common outputs |
|---|---|---|---|
| Autoregressive models | Generate a sequence one step at a time, conditioning each prediction on prior context. | GPT-style language models, PixelCNN | Text, code, audio, images |
| Variational autoencoders (VAEs) | Encode inputs into a probabilistic latent space and decode sampled latent vectors into new outputs. | Vanilla VAE, conditional VAE | Images, structured latent representations, synthetic records |
| Generative adversarial networks (GANs) | Train a generator against a discriminator so generated samples become harder to distinguish from training data. | StyleGAN, CycleGAN | Images, image-to-image translations |
| Diffusion models | Learn to reverse a gradual noising process and sample new outputs through iterative denoising. | DDPM, Stable Diffusion | Images, video, audio |
Generative vs discriminative models
| Dimension | Generative model | Discriminative model |
|---|---|---|
| Primary goal | Learn data patterns well enough to generate new samples. | Predict a label, score, or output from an input. |
| Typical output | New text, image, audio, code, or synthetic record. | Class label, probability, ranking score, or regression value. |
| Example task | Draft a support response or create an image from a prompt. | Classify a support ticket or detect whether an image contains a defect. |
| Evaluation focus | Quality, diversity, factuality, controllability, and safety. | Accuracy, precision, recall, calibration, and task-specific error costs. |
Use a generative model when the application must create or transform content. Use a discriminative model when the application primarily needs a prediction. Many production systems combine both: a discriminative classifier can route or validate output from a generative model.
Related terms
Explore adjacent ideas in the knowledge graph.
Related
Comparisons, tools, and models that connect to this idea.