Poisson Flow Generative Models (PFGMs)

• Poisson Flow Generative Models (PFGMs) are physics-inspired generative frameworks that model data with deterministic ODE flows in an augmented space using principles of electrostatics.
• PFGMs interpolate between normalizing flows and diffusion models by leveraging augmented dimensions and heavy-tailed kernels to balance robustness and sample fidelity.
• Extensions like PFGM++, Poisson Flow Consistency Models, and conditional adaptations enable efficient single-step and inverse problem sampling with state-of-the-art performance.

Poisson Flow Generative Models (PFGMs) are a class of physics-inspired generative models that cast data generation as a deterministic ordinary differential equation (ODE) flow in an augmented space, using principles from electrostatics and the Poisson equation. PFGMs provide a framework that interpolates between normalizing flows and diffusion models while introducing new tradeoffs in robustness, sampling efficiency, and sample quality. The PFGM family includes PFGM, its high-dimensional generalization PFGM++, distilled variants such as Poisson Flow Consistency Models (PFCM), and conditional/posterior-sampling adaptations for inverse problems.

1. Mathematical Foundations and Augmented Space Construction

PFGMs embed an $N$-dimensional data distribution $p(x)$, $x\in\mathbb{R}^N$, as a distribution of positive "charges" on the $z=0$ hyperplane in an $(N+1)$- or $(N+D)$-dimensional augmented space $\tilde{x} = (x, z)$. The framework views each data point as generating an electrostatic field via the solution of the Poisson equation in the higher-dimensional space. Specifically, the field obeys

$\nabla^2_{(x,z)} \Phi(x,z) = -p(x)\delta(z)$

with homogeneous (zero) Dirichlet boundary conditions at infinity. The resulting potential $\Phi(x,z)$ induces an "electric field" $E(x,z) = -\nabla_{(x,z)}\Phi(x,z)$, which governs a deterministic flow.

Sampling is recast as solving an ODE: $\frac{d\tilde{x}}{dt} = E(\tilde{x})$ starting from a tractable uniform distribution on a high-dimensional hemisphere (large $z$), and following field lines toward $z=0$, where the distribution on $x$ converges to $p(x)$. The reverse process integrates backward from the high-entropy prior to the data manifold, producing samples with the marginal $p(x)$ as $z\to0$ (Xu et al., 2022, Hein et al., 2024, Xu et al., 2023).

PFGM++ generalizes this construction to $D>1$ auxiliary dimensions: $\tilde{x} = (x, z) \in \mathbb{R}^{N+D}$. The rotational symmetry in $z$ reduces the dynamics to dependence solely on $r = \|z\|_2$.

2. Poisson Field, Perturbation Kernels, and Learning Objectives

The electric field at any point $\tilde{x} = (x, r)$ is given by

$$\mathbf{E}(\tilde{x}) = \frac{1}{S_{N+D-1}(1)} \int \frac{\tilde{x} - (y, 0)}{\|\tilde{x} - (y, 0)\|^{N+D}}\,p(y)\,dy$$

where $S_{N+D-1}(1)$ is the surface area of the unit $(N+D-1)$-sphere.

For practical learning, direct calculation of $\mathbf{E}(\tilde{x})$ is avoided; instead, a regression objective matches a trainable neural network $f_\theta$ to the direction of the field using randomly perturbed datapoints: $$\mathcal{L}(\theta) = \mathbb{E}_{r\sim p(r),\ y\sim p(y),\ x\sim p_r(x|y)} \left\| f_\theta(x, r) - \frac{x-y}{r/\sqrt{D}} \right\|^2$$ Here, $p_r(x|y)\propto ( \|x-y\|^2 + r^2 )^{-(N+D)/2}$ is the characteristic heavy-tailed Coulomb kernel and $D$ parametrizes the auxiliary dimension (Xu et al., 2023, Hein et al., 2023, Hein et al., 2024).

This loss directly estimates the field direction up to a scale, allowing unbiased, single-sample-per-step gradient estimates, eliminating the large-batch requirement of the original PFGM (Xu et al., 2022).

3. ODE Sampling and the Role of Augmentation Dimension $D$

Sampling in PFGM/PFGM++ is deterministic and proceeds by integrating the ODE

$$\frac{dx}{dr} = \frac{\mathbf{E}(x,r)_x}{\mathbf{E}(x,r)_r}$$

where $r = \|z\|$ is the augmented radius. When $D$ is large, the kernel $p_r(x|y)$ becomes increasingly Gaussian, and the formalism converges to score-based diffusion models (as shown by the limiting relation $r = \sigma\sqrt{D}$ with $\sigma$ the Gaussian noise level).

The parameter $D$ induces a continuum between the original Poisson flow ($D=1$), which is maximally robust due to the heavy kernel tails, and diffusion models ($D\to\infty$), which are more rigid but may be less robust to modeling error and step-size choices. Small $D$ broadens the support of $p_r(x|y)$, increasing robustness at the expense of sample quality saturation; large $D$ sharply localizes the kernel, improving asymptotic fidelity but with reduced error tolerance (Xu et al., 2023, Hein et al., 2024).

Crucially, after aligning $r = \sigma\sqrt{D}$, one can reuse EDM/diffusion-model noise schedules and hyperparameters, enabling straightforward zero-shot transfer across the $D$-axis (Xu et al., 2023).

4. Neural Network Architectures and Implementation Details

PFGM and its descendants typically use U-Net architectures with minor adaptations:

• The input is the data $x\in\mathbb{R}^N$ together with the scalar $z$ (or $r$), encoded as an extra channel or sinusoidal embedding, analogous to time-embedding in diffusion models.
• The output is a vector in $\mathbb{R}^{N+D}$, pooled if necessary to extract the $r$-component.
• Residual blocks use GroupNorm, SiLU (Swish) activations, and convolution, following best practices in score-based generative modeling (Xu et al., 2022, Xu et al., 2023).

Training uses Adam or similar optimizers, random perturbations at various radii or noise levels, and often exponential moving average of parameters for improved stability.

Sampling can be performed by Euler or second-order (Heun) ODE solvers, with NFE as low as 20–50 yielding acceptable sample quality, far fewer than in standard SDE/diffusion samplers (Xu et al., 2022, Xu et al., 2023).

5. Extensions, Posterior Sampling, and Single-Step Generative Inference

The PFGM family has been extended in several directions:

• PFGM++: Generalizes PFGM to arbitrary $D$; provides an unbiased, perturbation-based training objective; analytically interpolates between Poisson and Gaussian/diffusion dynamics (Xu et al., 2023).
• Poisson Flow Consistency Models (PFCM): Distilled, single-step variants enable direct mapping from noise to data in one forward pass, analogous to consistency distillation in diffusion (Hein et al., 2024).
• Conditional and Posterior Sampling: For inverse problems such as denoising or sparse-view CT, "hijacking" places a measured or low-quality input at an intermediate step in the ODE trajectory, proceeding with the learned flow or a single-step distillation. Heavy-tailed kernels (small $D$) contribute to robustness under this distribution mismatch (Hein et al., 2023, Hein et al., 2023, Hein et al., 2024, Fang et al., 6 Jun 2025).

These approaches facilitate efficient sampling (NFE = 1 in posterior applications) and strong empirical performance across low-dose and photon-counting CT denoising, sparse-view reconstruction, and supervised/unsupervised learning scenarios.

6. Empirical Benchmarking and Tradeoffs

PFGM, PFGM++, and descendants achieve state-of-the-art FID, Inception, and perceptual fidelity metrics in unconditional image synthesis and inverse imaging problems. Representative results include:

• CIFAR-10: PFGM achieves Inception 9.68 and FID 2.35 with a DDPM++ deep backbone (best among invertible flows to date); multi-step PFGM matches SDE samplers at 10–20$\times$ faster inference (Xu et al., 2022, Xu et al., 2023).
• PFGM++ on CIFAR-10/FFHQ-$64^2$: FID 1.91/2.43 for $D=2048/128$, exceeding EDM/diffusion benchmarks (Xu et al., 2023).
• Medical CT denoising (Mayo dataset): Single-step posterior-sampler PPFM ($D=64$ or $128$) achieves LPIPS down to 0.010, matching or surpassing supervised and consistency-distilled diffusion baselines with NFE=1 (Hein et al., 2023, Hein et al., 2023).
• Sparse-view CT reconstruction (ResPF): Outperforms FBPConvNet, SwinIR, EDM in SSIM, PSNR, and LPIPS while requiring fewer function evaluations (Fang et al., 6 Jun 2025).

A selection of empirical results for denoising and reconstruction tasks is summarized below:

Model/Setting	SSIM ↑	PSNR ↑	LPIPS ↓	NFE
PFGM, CIFAR-10 (uncond.)	—	—	—	110
PFGM++, CIFAR-10 $D=128$	—	—	1.92	35
Mayo CT, PPFM $D=64$	0.97	45.4	0.010	1
Mayo CT, Consistency Model	0.96	43.1	0.013	1
ResPF, Sparse CT	0.9639	39.80	0.0630	18

Smaller $D$ enhances robustness under large step size, model misspecification, and data mismatch at the cost of slightly increased trajectory diffusion; larger $D$ (diffusion limit) can be brittle under such conditions and less suited to single-step or "hijacked" inference (Xu et al., 2023, Hein et al., 2023, Hein et al., 2024).

7. Theoretical Insights and Connections to Physics-Inspired Generative Modeling

PFGM belongs to the broader GenPhys family of generative models, which establish generative flows via invertible mappings of linear partial differential equations (PDEs) derived from physics. The Poisson equation leads to a deterministic ODE flow, in contrast to the stochastic differential equations underpinning diffusion models (Liu et al., 2023). GenPhys encompasses PFGM, diffusion models, and generalizations to other physical PDEs (e.g., "Yukawa Generative Models" from screened Poisson equations).

PFGMs are distinguished by:

• Deterministic, non-stochastic generative flows.
• Use of a Cauchy or heavy-tailed prior in the augmented space.
• Exact likelihoods and the absence of blow-up in the learned vector field near $z\to0$.
• Accelerated burn-in and rapid progression to high-probability regions due to long-range Coulombic fields.

A key theoretical finding is that the parameter $D$ in PFGM++ explicitly controls the trade-off between robustness (to discretization, model, and step-size errors) and sample-quality rigidity, giving practitioners a degree of tuning unavailable in classical diffusion frameworks (Xu et al., 2023, Hein et al., 2024).

---

PFGMs, their generalizations, and distilled variants establish a powerful framework for generative modeling that unifies and extends over classical diffusion models. Their deterministic ODE-based flows, electrostatic interpretations, and empirical strengths in robustness and sampling efficiency make them foundational in contemporary physics-inspired generative modeling (Xu et al., 2022, Xu et al., 2023, Hein et al., 2024, Hein et al., 2023, Hein et al., 2023, Fang et al., 6 Jun 2025, Liu et al., 2023, Hein et al., 2024).