The Principles of Diffusion Models

TL;DR

A ~470-page monograph that unifies the three historically separate framings of diffusion models — variational (DDPM), score-based (NCSN / Score SDE), and flow-based (Neural ODE / Flow Matching) — under a single backbone: a time-dependent velocity field whose flow transports a simple prior to the data, with sampling reducing to solving a differential equation. The authors argue that the various training targets (noise, clean data, score, velocity) are algebraically interchangeable, with differences mattering only for implementation and stability rather than fundamental capability. On this foundation the monograph treats guidance for controllable generation, efficient numerical solvers, and diffusion-motivated flow-map models that learn direct mappings between arbitrary times. It is positioned as a reference text that consolidates work scattered across DDPM, Score SDE, Flow Matching, consistency models, and rectified flow into one mathematical scaffold.

Key claims

The variational view (VAE → DPM → DDPM), the score-based view (EBM → NCSN → Score SDE), and the flow-based view (NF → Neural ODE → Flow Matching) converge on the same object: a time-dependent velocity field whose flow connects a simple prior to the data distribution [Fig. 1, §A–B].
The Probability Flow ODE governs the deterministic sampling trajectory shared across all three views; consistency of the forward/reverse processes is guaranteed by the Fokker-Planck equation [§B].
The standard training targets — predicting noise, predicting clean data $x_0$ , predicting the score $\nabla_x \log p_t(x)$ , predicting velocity — are algebraically interchangeable parameterizations of the same underlying object; their differences are matters of variance and numerical stability rather than expressive power [§B, Part B].
Controllable generation via guidance (classifier guidance, classifier-free guidance, and variants) is presented as a separate axis on top of the unified velocity-field framework, with the guidance correction interpreted as a velocity modification rather than a sampler hack [Part C].
Efficient numerical solvers (higher-order ODE/SDE solvers, exponential integrators) are framed as drop-in replacements for the underlying ODE integration, independent of which of the three views one trains under [Part C].
Flow-map models — which learn direct mappings between arbitrary time pairs $(s, t)$ rather than iterative steps — are presented as the natural extension of the unified framework into the few-step / one-step regime, encompassing consistency models, rectified flow, and shortcut models as instances of the same object [Part D].
The monograph is structured in four parts: (A) introduction to deep generative modeling and the three lineages, (B) core perspectives unifying them, (C) guidance and efficient sampling, (D) flow-map models [Fig. 1, §0].

Method

This is a monograph, not an experimental paper. The “method” is the unification scaffold itself: start from a forward process that gradually corrupts data into noise, define the reverse process as a learned time-dependent velocity field, derive each of the three classical formulations (variational evidence lower bound, denoising score matching, flow matching) as different parameterizations of the same field, and then treat guidance, sampling, and flow-map distillation as transformations on that shared object. The authors trace the historical lineage of each view and show how the apparently distinct training losses collapse to interchangeable forms once expressed in terms of the velocity field and its associated Probability Flow ODE.

Results

No empirical results — this is a reference monograph. The contribution is pedagogical and structural: a single notational and conceptual framework that subsumes DDPM, Score SDE, Flow Matching, classifier-free guidance, DDIM/DPM-Solver-style samplers, and consistency / rectified-flow / shortcut models. Authorship signals seriousness — Yang Song (Score SDE, consistency models), Stefano Ermon (denoising score matching), Dongjun Kim, Yuki Mitsufuji, Chieh-Hsin Lai. Hugging Face listing reports 62 upvotes; ~470 pages.

Why it’s interesting

This is the canonical reference text for the framework that virtually every paper on the wiki tagged Diffusion Distillation, Classifier-Free Guidance variants, or Pixel-space diffusion implicitly assumes. The unification claim — that variational, score-based, and flow-based diffusion are the same object under different parameterizations — is directly what makes the wiki’s distillation thread coherent: TDM-R1: Reinforcing Few-Step Diffusion Models with Non-Differentiable Reward‘s unbiased-per-step-reward argument for deterministic ODE trajectories, ArcFlow: Unleashing 2-Step Text-to-Image Generation via High-Precision Non-Linear Flow Distillation‘s analytic mixture-of-momentum velocity, Transition Matching Distillation for Fast Video Generation‘s flow-head architecture, and Test-time scaling of diffusions with flow maps‘s flow-map test-time scaling all live inside Part D of this monograph. The guidance treatment in Part C is the natural backdrop for the Classifier-Free Guidance variants cluster’s recent reframings (SoFlow: Solution Flow Models for One-Step Generative Modeling folding CFG into training; Normalized Attention Guidance: Universal Negative Guidance for Diffusion Models moving extrapolation into attention space; Continuous Adversarial Flow Models arguing CFG patches generalization miscalibration). For a new team member, this is the single best on-ramp.