Diffusion Model
A diffusion model learns to reverse a gradual noising process. Sampling starts from noise and iteratively denoises toward the data distribution.
Definition
Forward diffusion adds Gaussian noise until the sample is ~N(0, I). A neural net (often a U-Net or DiT) predicts the noise or the score ∇ log p_t(x). Reverse sampling integrates that field. DDPM (Ho et al.) and score-based SDE (Song et al.) are the two standard views; they are closely related.
Why it matters
Diffusion displaced GANs for high-fidelity image, audio, and video generation and is the backbone of Stable Diffusion / Sora-class systems. The likelihood-ish training is stabler than adversarial games.
Case
DDPM on CIFAR produced samples that finally matched GAN FID without a discriminator. The cost is many network evaluations per sample — the reason latent diffusion and distillation exist.
How to read it
Separate the noising schedule, the predictor target (ε, v, or x₀), and the sampler (DDPM, DDIM, Heun). Those three knobs change quality more than swapping the U-Net’s brand name.
Key papers
Landmark references with real publisher or arXiv links — not token-overlap noise.
- Deep Unsupervised Learning using Nonequilibrium ThermodynamicsSohl-Dickstein et al. · 2015 · ICMLDiffusion as a generative process.
- Denoising Diffusion Probabilistic ModelsHo, Jain, Abbeel · 2020 · NeurIPSDDPM — the modern recipe.