A Unified View of Drifting and Score-Based Models

TL;DR

Drifting models (Kaiming-lab style one-step generators that follow a kernel-induced mean-shift field) are shown to be exactly a score-matching objective on kernel-smoothed distributions. For Gaussian kernels, the population mean-shift field equals the score difference between Gaussian-smoothed data and model distributions — a direct consequence of Tweedie’s formula — making Gaussian-kernel drifting a reverse-Fisher (model-weighted) variant of Distribution Matching Distillation, except the score is realized nonparametrically from samples rather than via a pretrained diffusion teacher. For the Laplace kernel actually used in the original drifting paper, the authors derive an exact preconditioned-score decomposition and prove polynomial-decay error bounds in the low-temperature and high-dimensional regimes.

Key claims

For Gaussian kernels with bandwidth σ, the mean-shift direction equals σ² times the score difference between the Gaussian-smoothed data and model distributions, via Tweedie’s formula (kernel barycenter = Bayes-optimal denoiser; mean-shift = denoising residual) [§4.1, Theorem 1].
Gaussian-kernel drifting and DMD share the same score-mismatch transport direction under reverse-Fisher (model-weighted) weighting — drifting realizes the score nonparametrically from kernel neighborhoods, DMD via a pretrained diffusion teacher [§4.1, Connection-to-DMD subsection].
Drifting fits a transport–then–projection template: a stop-gradient one-step transport produces a target sample cloud, then the next generator is fit by regression onto that cloud — formally equivalent in objective value to the squared-norm functional of the drift field under the model [§3.1, Eq. 4].
Identifiability holds for Gaussian-kernel drifting: zero mean-shift field ⇒ matching smoothed scores ⇒ matching smoothed densities ⇒ matching data distributions (by injectivity of Gaussian smoothing) [§4.1, Proposition 1].
For general radial kernels k(x,y)=ψ(‖x−y‖²), mean-shift exactly decomposes into a preconditioned smoothed-score term plus a scalar-vector covariance residual capturing distance-direction coupling under the local kernel-reweighted law [§4.2, Theorem 2].
For the Laplace kernel (default in drifting implementations), the off-score residual is a radius-weighted average of the tangential direction and vanishes only when radius and tangential direction are uncorrelated under the local reweighted law [§4.2, post-Theorem-2 discussion].
For Laplace, polynomial-decay error bounds hold in two regimes: (1) low temperature (small bandwidth) — kernel localizes and the population minimizer matches the kernel-smoothed score up to error polynomial in σ; (2) high dimension — vector-field, update, and minimizer alignment all hold with discrepancy decaying polynomially in d [§4.3, Theorems 4, 5].
The drifting objective is in reverse Fisher form (averaged under the model, not the data), which makes it susceptible to mode dropping early in training where the model has zero mass on a true mode [§4.1, post-Theorem-1 mode-dropping remark].

Method

The paper is a theoretical bridge, not an algorithm. The setup is the fixed-point regression template of the original drifting work: a pushforward generator G induces a model distribution p, a drift field D(x; p, q) is designed to vanish at equilibrium with the data q, and training is the stop-gradient regression

L(θ) = E_{z∼ρ₀}[ ‖G_θ(z) − sg(G_θ(z) + D(G_θ(z); p, q))‖² ]

which equals ‖D‖²_p in objective value (Eq. 4). Two realizations of D are studied:

Mean-shift: D_meanshift(x) = (∫k(x,y) y dq(y))/(∫k(x,y) dq(y)) − (analogous for p) — the kernel-weighted barycenter difference.
Kernel-induced score: D_score(x) = ∇log(kq)(x) − ∇log(kp)(x) — the score difference between kernel-smoothed densities.

The main result is identifying these two fields. For Gaussian k_σ(x,y) = exp(−‖x−y‖²/2σ²) the identity is exact: D_meanshift = σ² · D_score (Theorem 1, proved by Tweedie’s formula). For general radial k(x,y) = ψ(‖x−y‖²) the identity becomes a preconditioned decomposition (Theorem 2): mean-shift = α(x) · (preconditioned smoothed score) + scalar-vector covariance residual, where α(x) is a kernel-dependent scaling and the residual captures how the average displacement under the local reweighted law varies with radius. The off-score component (perpendicular to the score direction) is then bounded for the Laplace kernel under low-temperature and high-dimension limits.

The paper also formalizes DMD as a special case: DMD’s gradient direction reduces to a score-mismatch under reverse-Fisher weighting (Eq. 2), and fits the same transport–then–projection template (Eq. 8), with the only differences being (a) the score signal comes from a pretrained teacher rather than kernel neighborhoods and (b) DMD naturally includes a noise level whereas mean-shift drifting evaluates on clean model samples (or, equivalently after Tweedie, on Gaussian- noised samples at the kernel bandwidth).

Results

This is a theoretical paper; the experimental content is validation plots. The headline empirical claims:

The predicted Gaussian-kernel mean-shift = score-mismatch identity is visually confirmed: the two vector fields coincide up to the σ² scaling [Fig. 1(c,d)].
The predicted polynomial decay rate of the Laplace off-score error in dimension d is empirically reproduced [§4.3, validation experiments referenced].
Gaussian-kernel and Laplace-kernel drifting yield “generally comparable generation quality” in the regimes the theory predicts [§1, summary paragraph].

There are no new FID numbers or generative-quality wins claimed against external baselines — the contribution is a unification, not a new generator.

Why it’s interesting

This is the theoretical glue between two threads on the wiki that looked unrelated. Generative Modeling via Drifting presented drifting as a fundamentally different formulation from diffusion / flow models — “fixes neither the noising process nor the flow ODE, lets SGD itself be the sampler.” This paper says: actually, Gaussian-kernel drifting is a score-matching objective on smoothed distributions, just realized teacher-free via Tweedie/mean-shift instead of via a pretrained diffusion teacher. That collapses the “new paradigm” framing into a known one (DMD-with-different-score- source) for the Gaussian case, and quantifies how far the actually- deployed Laplace kernel deviates from this in two limits.

It also clarifies a confusion in Diffusion Distillation: the reverse-Fisher / mode-seeking weighting that explains DMD’s diversity problems (Diversity-Preserved Distribution Matching Distillation for Fast Visual Synthesis attacks this with role-separation) is the same weighting drifting inherits, predicting that drifting should have analogous mode-dropping behavior early in training. The paper flags this directly. For Luma’s few-step generation interests, the practical implication is that Drifting’s FID 1.54 result is not compelling evidence for a fundamentally new paradigm being needed — the same score-mismatch direction underlies DMD-class teacher distillation, and the open question becomes whether the nonparametric/teacher-free realization buys anything beyond avoiding the teacher cost. The paper is honest that the kernel-bandwidth and sample-density requirements may eat that win in high-dimensional settings.