Generative Models Basics — Density/Energy Models, GANs, Autoregressive, Diffusion | MIT 6.7960 Lecture 14

🌏 中文版

Video status: Videos included. Source details

Source version: based on MIT 6.7960 Fall 2024 OCW. Videos, slides, and assignments are public at MIT OCW. This lecture is taught by Phillip Isola; the required reading is Generative Models.

The ultimate goal of generative models

Every generative model does the same thing: learn the data distribution p(x) so that sampling from it produces samples "of the same kind but brand new". They differ only in how they represent it, how they train it, how they sample from it.

Density models: model probability directly

The most intuitive is the explicit density model: parameterize p_θ(x) and maximize the log-likelihood log p_θ(x) of the training data.

• Autoregressive models: factorize p(x) as p(x_1)·p(x_2|x_1)·…·p(x_n|x_<n), predicting the next element step by step. PixelCNN generates images pixel by pixel this way. Upside: likelihood is exact and training is stable. Downside: slow generation (must be sequential).
• Normalizing flows: a chain of invertible transforms mapping a simple distribution to a complex one, with likelihood still exactly computable.

Energy models: define "looks real" first, then sample

An energy-based model (EBM) does not give a probability directly; it defines an energy function E_θ(x) such that

p_θ(x) = exp(−E_θ(x)) / Z_θ        (Z_θ is an intractable normalization constant)

Low E_θ(x) = dense data region = high probability. The catch: Z_θ is usually a high-dimensional integral you cannot compute. So training and sampling split:

• Training: use contrastive divergence and friends — only require "real samples have low energy, generated samples have high energy", dodging Z_θ.
• Sampling: use Langevin dynamics / MCMC, walk along −∇E toward low energy, adding noise so you do not get stuck.

EBMs are appealing for their expressiveness and freedom from a tractable density form; the cost is slow sampling and brittle training.

GAN: let a discriminator force realistic samples

A GAN uses an adversarial game to dodge explicit density: a generator G fakes samples, a discriminator D tells real from fake, in a zero-sum tug-of-war.

min_G max_D  E[log D(x)] + E[log(1 − D(G(z)))]

Once training stabilizes, G's outputs are visually striking (GAN's strength). But the cost is fragile training: mode collapse (only a few modes generated), vanishing gradients, hyperparameter sensitivity. Later WGAN, spectral norm, and GAN-stabilization tricks patch these holes.

Diffusion: add noise step by step, then learn to denoise

Diffusion models (DDPM) are the strongest image generators of recent years, with an elegant idea:

1. Forward process: gradually add Gaussian noise to a real sample; after T steps it becomes pure noise.
2. Reverse process: train a network ε_θ(x_t, t) to predict "how much noise was added this step" — equivalently, learn to denoise.
3. Sampling: start from pure noise, repeatedly predict and subtract noise, restoring a sharp sample step by step.

This dodges the intractable normalization constant of maximum likelihood; the training objective is just a simple "predict the noise" MSE, yet it generates extremely high-quality images. The cost is many sampling steps (though DDIM and friends later sped it up).

A PyTorch-style training objective (simplified):

def diffusion_loss(net, x0, t, sqrt_alphas_cumprod, sqrt_1_m_alphas):
    noise = torch.randn_like(x0)
    x_t = sqrt_alphas_cumprod[t] * x0 + sqrt_1_m_alphas[t] * noise  # add noise
    pred_noise = net(x_t, t)
    return F.mse_loss(pred_noise, noise)                            # predict noise

How to choose

• Need exact likelihood, want stability: autoregressive / flow.
• Need极致 image fidelity: GAN (but hard to train) or diffusion (stable and high quality).
• Need to express arbitrarily complex distributions, free of density form: energy models.
• Discrete sequences like text: autoregressive (today's LLMs are exactly this).

The next batch (L15, L16) goes deep on VAEs and conditional generation (text-to-image, image-to-text), connecting generative models back to the need for control.

Course video sources

Rechecked against the live MIT OCW Fall 2024 gallery on 2026-10-10: the lecture number matches and the video is public and embeddable. Lecture numbering and topics follow OCW.

Original videos: MIT 6.7960 Fall 2024 — Lec 14. Generative Models: Basics

Course and recording entries:

• MIT OCW — Lec 14. Generative Models: Basics
• MIT 6.7960 Fall 2024 — official lecture video gallery

Checked: 2026-10-10.

Content check: verified against the video transcript (2026-10-10): read the Lec 14 transcript (Phillip Isola, 1:21:18): the video previews autoregressive, diffusion and GAN models, and its body covers probability densities and maximum likelihood (about 16%–50%), energy models (about 52%–70%), autoregressive models (about 70%–77%), diffusion (about 85%–95%) and GANs (about 96%), matching the article; nothing needed correcting.

Update Log

• 2026-10-10: Added explicit video status and checked recording sources and access notes.
• 2026-10-10: Rechecked video status. Live-checked the official course page: lecture numbers and recording links match and the videos are public, so the status is now “Videos included.”
• 2026-10-10: Checked the video content against its transcript. The video topic matches the article; no correction needed.

References

• MIT 6.7960 OCW (Fall 2024): course home
• Goodfellow et al., Generative Adversarial Networks (GAN): arXiv:1406.2661
• Ho et al., Denoising Diffusion Probabilistic Models (DDPM): arXiv:2006.11239
• Van den Oord et al., Conditional Image Generation with PixelCNN: arXiv:1606.05328