Generalized Poisson Flow (GPFlow) Overview

• Generalized Poisson Flow (GPFlow) is a family of generative constructions that uses Poisson dynamics in augmented spaces to model fixed-dimensional images and variable-length sequences.
• It replaces Gaussian diffusion with Poisson-based transport laws, offering a tunable trade-off between robustness and generation speed via parameters like the augmentation dimension D.
• Applications range from high-fidelity image synthesis and low-dose CT denoising to state-of-the-art variable-length protein design, achieving competitive results with fewer computational steps.

Generalized Poisson Flow (GPFlow) denotes a family of generative constructions organized around Poisson-type dynamics, but the term is used in two closely related senses in the recent literature. In the fixed-dimensional image-generation line, it refers to generalizations of Poisson Flow Generative Models (PFGM), where data are treated as electric charges in an augmented space and generation follows deterministic ODEs induced by a Poisson field (Xu et al., 2022). In the variable-length protein-design line, GPFlow is a specific framework that learns the rate function of an inhomogeneous generalized Poisson process and couples length insertions to within-length continuous, discrete, or Riemannian dynamics, thereby modeling joint distributions over both length and content (Cheng et al., 10 Jul 2026). The common thread is the replacement of a purely Gaussian, fixed-dimensional forward corruption process by transport laws derived from Poisson equations, generalized Poisson processes, or their limiting relations.

1. Terminological scope and historical development

The modern point of departure is the 2022 formulation of PFGM, which proposes a generative model that maps a uniform distribution on a high-dimensional hemisphere into an arbitrary data distribution by interpreting data points as electrical charges on the $z=0$ hyperplane in an augmented space and following electric field lines generated by the solution to a Poisson equation (Xu et al., 2022). In that formulation, the additional coordinate $z$ is not an auxiliary bookkeeping device: it anchors the backward ODE, and the sampling trajectory terminates when $z$ reaches zero.

Subsequent work generalized this construction in two directions. First, PFGM++ extends the augmented space from one extra coordinate to $D$ extra coordinates and treats $D$ as a design parameter. The 2024 Poisson Flow Consistency Models paper explicitly frames this as a broader family of Poisson-type flows parameterized by $D$, and shows that consistency distillation can be ported from diffusion probability-flow ODEs to PFGM++ by a change of variables and an updated noise distribution (Hein et al., 2024). Second, the 2025 Poisson Flow Joint Model for multiphase CT treats $D$ as a task-specific generation-path parameter and uses PFGM++ as the backbone for conditional joint modeling of routine-dose and low-dose multiphase CT data (Ge et al., 5 Feb 2025).

A common misconception is to treat all usages of “GPFlow” as simple renamings of PFGM++. The literature does not support that identification. In the image-generation line, “generalized” refers to extensions of electrostatic Poisson flow in augmented Euclidean space, especially through the augmentation dimension $D$. In the 2026 protein-design line, “Generalized Poisson Flow” is the name of a variable-length probability-flow framework whose central object is an inhomogeneous generalized Poisson process over sequence length rather than an electrostatic field over a fixed-dimensional ambient space (Cheng et al., 10 Jul 2026). This suggests that GPFlow is best understood as a family resemblance rather than a single algorithm.

2. Electrostatic Poisson flow in augmented space

In the original PFGM construction, data live in $\mathbb{R}^N$ and are embedded into an augmented space $\mathbb{R}^{N+1}$ through

$z$0

with the data distribution represented as a charge sheet on the hyperplane $z$1: $z$2 The corresponding Poisson equation in the augmented space is

$z$3

and the electric field is the negative gradient of the potential,

$z$4

The model’s base distribution is a uniform distribution on the upper hemisphere $z$5 of radius $z$6 in $z$7, and a central theorem states that, under compact-support and regularity assumptions, backward flow from a uniform hemisphere recovers the data distribution in the limit $z$8 (Xu et al., 2022).

The associated flow is deterministic. In augmented coordinates $z$9, the forward and backward ODEs are

$z$0

A key structural fact is rescalability: replacing $z$1 by $z$2 for any positive scalar field $z$3 preserves the same trajectories up to time reparameterization. PFGM exploits this by learning only the normalized field direction,

$z$4

rather than the full field magnitude. This makes the learned object a direction field whose integral curves coincide with those of the physical Poisson field.

Sampling is performed with a backward ODE anchored by the physically meaningful coordinate $z$5. After reparameterizing by $z$6 and then by $z$7, the model uses

$z$8

so that $z$9 decays exponentially and the stopping criterion is simply $D$0. This anchoring is central to the original claim that samples “hit the unaugmented data manifold when the $D$1 reaches zero” (Xu et al., 2022).

The significance of this formulation is empirical as well as geometric. On CIFAR-10, PFGM reports an Inception score of $D$2 and an FID score of $D$3, described as current state-of-the-art performance among normalizing flow models, while also performing on par with state-of-the-art SDE approaches and offering $D$4 to $D$5 acceleration on image generation tasks. The same paper also reports tolerance to estimation errors on a weaker network architecture and robustness to the step size in the Euler method (Xu et al., 2022).

3. PFGM++, augmentation dimension $D$6, and the diffusion limit

PFGM++ generalizes the augmented-space construction by embedding data into $D$7 rather than $D$8. With

$D$9

the data are again treated as electric charges on the hyperplane $D$0, and the electrostatic field is

$D$1

Because of rotational symmetry in the augmented coordinates, it is sufficient to track the data component $D$2 and the radius

$D$3

The Poisson flow ODE then reduces to

$D$4

which defines a bijective mapping between the data manifold at $D$5 and a noise hypersurface at $D$6 (Hein et al., 2024).

Training is recast through a perturbation kernel rather than direct field evaluation: $D$7 The objective is

$D$8

whose minimizer is

$D$9

Thus the network directly approximates the ODE vector field $D$0 up to the factor $D$1 (Hein et al., 2024).

The role of $D$2 is the decisive generalization lever. With the change of variables

$D$3

PFGM++ can be written in the same generic probability-flow ODE form used in diffusion models,

$D$4

with $D$5. The same line of work states that, with $D$6 and $D$7, the PFGM++ objective and sampling converge to EDM, while $D$8 recovers the original PFGM (Hein et al., 2024). The 2025 CT paper reiterates this as the central unification result: PFGM++ becomes PFGM for $D$9 and the diffusion model for $D$0 (Ge et al., 5 Feb 2025).

This parameterization gives GPFlow, in the fixed-dimensional sense, a continuum of behaviors rather than a binary choice between Poisson flow and diffusion. Small finite $D$1 yields non-Gaussian, heavy-tailed perturbations and different robustness properties; large $D$2 yields diffusion-like behavior. The literature repeatedly interprets this as a robustness–rigidity trade-off.

4. Consistency formulations and low-dose CT denoising

Poisson Flow Consistency Models (PFCM) combine PFGM++ with the consistency-model framework, yielding a family of Poisson-flow-based consistency models that support single-step or few-step sampling (Hein et al., 2024). The key formal move is to generalize consistency distillation from diffusion probability-flow ODEs to Poisson-flow ODEs by modifying both the solver update and the perturbation distribution. In place of Gaussian perturbations, PFCM use the PFGM++ heavy-tailed kernel, which in radial form satisfies

$D$3

The learned consistency function maps noisy states at scale $D$4 to the near-clean endpoint: $D$5 with boundary condition $D$6. In the conditional CT formulation, the low-dose image $D$7 is added as an extra input channel, and $D$8 directly approximates the posterior mapping associated with $D$9. Distillation uses a reverse PF ODE step

$D$0

and a consistency matching loss with LPIPS distance and an EMA teacher.

The CT application is explicitly described as a “task-specific” sampler that “hijacks” the generative process by replacing an intermediate state with the low-dose CT image. The paper also states that this introduces a severe mismatch because the noise characteristics of low-dose CT images are different from those of intermediate states in the Poisson flow process. The central empirical claim is that small $D$1 mitigates this mismatch: PFCM inherit from PFGM++ the ability to trade off robustness for rigidity via the hyperparameter $D$2, and the resulting low-$D$3 samplers are markedly more robust than standard consistency models in this conditional setting (Hein et al., 2024).

On the Mayo low-dose CT validation set, the reported mean $D$4 std metrics are: LDCT input, LPIPS $D$5, SSIM $D$6, PSNR $D$7; EDM at $D$8 NFE, LPIPS $D$9, SSIM $\mathbb{R}^N$0, PSNR $\mathbb{R}^N$1; CD at $\mathbb{R}^N$2 NFE, LPIPS $\mathbb{R}^N$3, SSIM $\mathbb{R}^N$4, PSNR $\mathbb{R}^N$5; PFGM++ with $\mathbb{R}^N$6 at $\mathbb{R}^N$7 NFE, LPIPS $\mathbb{R}^N$8, SSIM $\mathbb{R}^N$9, PSNR $\mathbb{R}^{N+1}$0; and PS-PFCM with $\mathbb{R}^{N+1}$1 at $\mathbb{R}^{N+1}$2 NFE, LPIPS $\mathbb{R}^{N+1}$3, SSIM $\mathbb{R}^{N+1}$4, PSNR $\mathbb{R}^{N+1}$5. The same study reports that the $\mathbb{R}^{N+1}$6-NFE sampler reduces CT validation processing from approximately $\mathbb{R}^{N+1}$7 hours to approximately $\mathbb{R}^{N+1}$8 minutes on a single A6000 GPU, and that unconditional PFCM on CIFAR-10 reaches FID $\mathbb{R}^{N+1}$9 at $z$00 NFE (Hein et al., 2024).

5. Joint conditional generalized Poisson flow for multiphase CT

The Poisson Flow Joint Model (PFJM) is a conditional, joint instantiation of PFGM++ for low-dose multiphase contrast-enhanced CT, where the data variable is the joint routine-dose triplet

$z$01

and the conditioning signal is the corresponding low-dose triplet

$z$02

The model treats the full triplet as a single datum in the generalized Poisson-flow framework, with effective data dimension $z$03, and learns the conditional mapping associated with

$z$04

rather than modeling each phase independently (Ge et al., 5 Feb 2025).

Training follows the PFGM++ perturbation design. A noise scale $z$05 is sampled, the augmented-space radius is set by

$z$06

a scalar perturbation $z$07 and a random direction $z$08 with $z$09 are drawn, and the perturbed joint image is

$z$10

The supervised conditional loss is

$z$11

This differs from the original PFGM++ objective by predicting the clean joint routine-dose image directly and by conditioning explicitly on the joint low-dose image.

Sampling adds a second conditioning mechanism at the path level. The initial state is set to the condition,

$z$12

and at each ODE step the current state is pulled toward the condition by

$z$13

The update direction is then computed from the network output, followed by a trapezoidal correction. The paper describes this as stronger than merely conditioning the network, because conditioning enters both the denoiser/vector field and the generation path itself.

PFJM treats $z$14 as a task-specific generation-path parameter and evaluates $z$15. The reported optimum is $z$16, with performance improving rapidly from $z$17 to $z$18 and then degrading gradually for larger $z$19. Across all tested $z$20 and phases, PFJM outperforms PFGM++, with average gains of $z$21 HU in MAE, $z$22 in SSIM, and $z$23 dB in PSNR (Ge et al., 5 Feb 2025).

The headline quantitative results for PFJM averaged over phases are MAE $z$24 HU, SSIM $z$25, and PSNR $z$26 dB. The reported FID over joint 3-channel images is $z$27, compared with $z$28 for PFGM++ with $z$29, $z$30 for DDIM with $z$31 steps, and $z$32 for PFGM. PFJM also uses $z$33 ODE steps, whereas the diffusion baselines use $z$34 and PFGM uses $z$35 (Ge et al., 5 Feb 2025). In this sense, PFJM is a concrete demonstration that generalized Poisson flow can support joint, strongly conditional medical-image generation with fewer deterministic solver steps than the compared diffusion baselines.

6. Variable-length GPFlow for protein design

In the 2026 protein-design framework, GPFlow is no longer a fixed-dimensional electrostatic flow. It is a variable-length generative model over

$z$36

where $z$37 is the current protein length and $z$38 is the component space for residues, such as $z$39 for coordinates, a categorical alphabet for amino acids, $z$40 for rotations, or $z$41 for torsion angles. Length evolves through a continuous-time pure birth process $z$42 with rate function $z$43, and content evolves through within-length generators $z$44 appropriate to the modality (Cheng et al., 10 Jul 2026).

For a fixed target length $z$45, GPFlow introduces a scheduler $z$46 and defines

$z$47

with completion time $z$48. Under this construction, $z$49 for $z$50, and $z$51 almost surely as $z$52. For an arbitrary target length distribution $z$53, the marginal rate is obtained by posterior averaging,

$z$54

Theorem 3.2 states that the generalized Poisson process driven by $z$55 has marginals $z$56 matching the posterior mixture over conditional paths and satisfies $z$57.

A central technical contribution is the simulation-free likelihood objective. For a realization with event times $z$58, the negative log-likelihood of the generalized Poisson trajectory is

$z$59

After marginalizing event times analytically, the rate-learning objective becomes

$z$60

This trains the marginal rate network $z$61 without simulating the event process during training.

The multimodal extension couples length insertions to within-length dynamics. For current state $z$62, insertions are represented by an insertion variable $z$63 and map $z$64, while the conditional jump measure is

$z$65

The corresponding marginal objects are

$z$66

$z$67

Training combines the rate loss with a reconstruction term for insertion sampling and a within-length flow-matching term,

$z$68

The associated theory is formulated at the level of general Markov generators. For a process with generator

$z$69

Theorem 4.1 gives the bound

$z$70

where

$z$71

For GPFlow, the jump-measure divergence decomposes as

$z$72

and the paper states that, at the global optimum where these terms vanish, the learned process recovers the joint multimodal distribution over variable-length proteins.

Empirically, GPFlow is evaluated on unconditional structure design, unconditional sequence design, structure-based motif scaffolding, sequence-based motif scaffolding, and peptide co-design. In unconditional structure design, it raises designability on PDB from $z$73 for Proteina (60M) to $z$74 for GPFlow (65M), and on AFDB from $z$75 to $z$76, while recovering the empirical length distribution with mean $z$77 versus empirical $z$78. In unconditional sequence design, it matches UniRef50 mean pLDDT more closely than DPLM, with a difference of $z$79 versus $z$80, while maintaining approximately $z$81 higher diversity than DPLM. In structure-based motif scaffolding, it ranks first on $z$82 of $z$83 tasks, with average unique successes $z$84 and total unique successes $z$85. In sequence-based motif scaffolding, it passes $z$86 tasks versus $z$87 for DPLM and $z$88 for EvoDiff. In peptide co-design, despite not using native length conditioning, it improves over PepFlow in AAR ($z$89 versus $z$90), RMSD ($z$91Å versus $z$92Å), designability ($z$93 versus $z$94), and diversity ($z$95 versus $z$96) (Cheng et al., 10 Jul 2026).

In this protein-design usage, GPFlow refers specifically to probability flows over variable-dimensional state spaces in which length emerges as the realized count of insertion events. That distinguishes it sharply from the fixed-dimensional electrostatic construction of PFGM and PFGM++, even though both lines of work use Poisson-type objects to define generative trajectories.