Langevin Dynamics and Score-Based Generative Models

The Score Function: A Compass Pointing Toward Likely Data

Every dataset, whether it's a collection of cat photos or measurements of protein shapes, can be thought of as samples drawn from some underlying probability density p(x). We rarely know p(x) explicitly, but we can still ask a more local question: if I nudge a point x slightly, which direction increases the probability fastest? The answer is the score function, defined as s(x) = ∇_x log p(x), the gradient of the log-density with respect to the data itself (not with respect to any parameters). Geometrically, the score is a vector field laid over the whole data space: at every point x it points in the direction of steepest ascent of probability, and its magnitude tells you how sharply probability is changing there. Near a mode of the distribution the score vectors converge inward from all sides, like water flowing downhill into a basin, while in the sparse, low-probability regions between clusters the score field carries points firmly back toward the nearest cluster. Working with the log-density rather than the density itself is not just a convenience — it also strips out the pesky normalizing constant Z that makes most real densities p(x) = exp(-E(x))/Z impossible to compute exactly, since log p(x) = -E(x) - log Z and the gradient with respect to x kills the constant term entirely.

Langevin Dynamics: Gradient Ascent Plus a Dose of Randomness

If the score always points toward higher probability, why not just follow it uphill, the way ordinary gradient ascent would? Because doing so would collapse every starting point onto the single tallest peak (the mode) of the distribution, destroying all the diversity that makes a distribution a distribution. Langevin dynamics fixes this by adding calibrated noise to every step, producing the update rule x_t+1 = x_t + (ε/2)∇_x log p(x_t) + √ε · z_t, where ε is a small step size and z_t is a fresh draw from a standard Gaussian, z_t ~ N(0, I). The first term is ordinary score-guided drift, pulling the point toward regions of higher probability; the second term is thermal-style jitter, constantly kicking the point sideways. This is precisely the discretized form of a stochastic differential equation originally used by Paul Langevin in 1908 to describe a pollen grain buffeted by water molecules (Brownian motion) while also being pulled by an external force. The remarkable mathematical fact, guaranteed by the Fokker-Planck equation, is that as ε → 0 and the number of steps grows large, the distribution of x_t converges exactly to p(x) itself — not to its mode, but to the full, correctly weighted shape of the distribution, with dense regions visited proportionally often and sparse regions visited rarely. The noise is not a bug tolerated for exploration; it is the mathematically essential ingredient that turns an optimizer into a sampler.

Learning the Score Without Ever Knowing the Normalizer

Langevin dynamics needs the score s(x) = ∇_x log p(x) at every point, but for real data like images, p(x) is never available in closed form. The elegant workaround is score matching: train a neural network s_θ(x) to directly approximate the score field, without ever estimating p(x) or its normalizing constant Z. The natural-looking objective, minimizing the expected squared error E_p(x)[||s_θ(x) - ∇_x log p(x)||²], is unusable as written because it still references the unknown ∇_x log p(x). Hyvarinen's original score matching result shows this objective is equivalent, up to a constant independent of θ, to E_p(x)[tr(∇_xs_θ(x)) + ½||s_θ(x)||²], which involves only the network's own Jacobian trace and norm, both computable from samples alone. In practice, modern systems favor an even simpler variant called denoising score matching: deliberately corrupt a clean data point x with Gaussian noise to get x̃ = x + σz, and train the network to predict s_θ(x̃) ≈ -z/σ, which can be shown to equal the true score of the noised distribution. This reframes score estimation as an ordinary denoising problem solvable with plain regression, sidestepping the normalizing constant entirely and making it trivial to train at massive scale.

Multiple Noise Scales: The Mathematical Core of Diffusion Models

A single, fixed noise level for Langevin dynamics runs into trouble in high dimensions: real data like images occupies a thin manifold within a vast pixel space, so the true data density is near zero almost everywhere, and the score is poorly estimated and wildly noisy far from the data. The fix, introduced by Song and Ermon, is to train the score network across a whole sequence of noise scales σ₁ > σ₂ > ... > σ_L, from very blurry and easy to score down to nearly clean and precise. Sampling then proceeds through annealed Langevin dynamics: start from pure noise, run several Langevin steps at the largest σ₁ to make coarse, confident progress across the smoothed-out landscape, then anneal down to σ₂, then σ₃, and so on, each stage refining detail while the earlier stages have already placed the sample in roughly the right neighborhood. This annealing scheme is mathematically the same idea that underlies denoising diffusion probabilistic models (DDPMs): a forward process gradually adds Gaussian noise to data across many timesteps until it becomes indistinguishable from pure noise, and a learned reverse process removes that noise step by step. Both formulations converge on the same continuous-time object, a score-guided stochastic differential equation, x_t+1 = x_t + (ε_t/2)∇_x log p_t(x_t) + √ε_t · z_t, run backward from noise to data at a schedule of shrinking noise levels σ_t. Every modern diffusion model that generates images, audio, or 3D shapes is, at its mathematical heart, exactly this: a neural score estimator paired with noisy, annealed Langevin steps.

Why Noise Helps Rather Than Hurts

It can feel counterintuitive that deliberately corrupting a search with randomness makes it better rather than worse, but the noise solves two distinct problems at once. First, it is what allows the sampler to represent an entire distribution instead of collapsing to a point estimate: without the √ε · z_t term, every trajectory started near the same basin of attraction would converge to that basin's single mode, and the diversity of outputs — different faces, different cats, different plausible completions — would vanish. Second, in the multi-scale setting, injecting noise during training (as in denoising score matching) forces the network to learn a smoothed, well-behaved score field even in regions where real data is sparse, because noisy versions of data points spread out to cover the space that clean data leaves empty. This is exactly why a Langevin-based sampler can start from a completely random image, wander through an initially uninformative score field, and still reliably arrive at a sharp, coherent sample: the annealing schedule guarantees that a useful score signal exists at every stage of the journey, from the coarsest noise level down to the finest.

Frequently asked questions

Try it live

Everything above runs in your browser — open Langevin Dynamics and Score-Based Generative Models and change the parameters while it is running. Nothing is installed, nothing is uploaded, the whole model lives in one tab.