Projected Conditional Flow Matching (PCFM)

• PCFM is a generative modeling framework that projects latent flow onto constraint manifolds for exact satisfaction of physical or data consistency constraints.
• It employs Gauss–Newton projections and relaxed corrections to enforce nonlinear equality constraints without retraining pretrained models.
• PCFM demonstrates superior performance in PDE-constrained simulations and MRI reconstruction, boosting sample quality and constraint fidelity.

Projected Conditional Flow Matching (PCFM) is a framework that enforces hard constraints—such as physical laws or data consistency—during zero-shot generative inference or unsupervised learning in normalizing flow frameworks. PCFM projects the latent flow evolution of a generative model onto a constraint manifold, ensuring exact satisfaction of arbitrary nonlinear constraints at final sample time, and is broadly applicable in scientific machine learning and inverse problems. PCFM has been shown to significantly improve constraint fidelity and sample quality in PDE-constrained generation as well as high-dimensional inverse imaging, all while requiring no retraining or architectural modifications of pretrained flow models (Utkarsh et al., 4 Jun 2025, Luo et al., 19 Dec 2025).

1. Mathematical Foundation of PCFM

PCFM operates within the continuous-time normalizing flow (CNF) paradigm, where a neural ODE with a learned vector field $v_\theta(x, t)$ transports a simple base distribution $\pi_0$ (e.g., isotropic Gaussian) to a target distribution $\pi_1$ via the ODE

$$\frac{dx(t)}{dt} = v_\theta(x(t), t), \quad x(0) \sim \pi_0, \quad x(1) \approx \pi_1.$$

Classical flow-matching loss is defined over pairs $(x_0, x_1)$ by interpolating along the OT path $x_t = (1-t)x_0 + t x_1$ and matching the learned velocity field to the optimal velocity $u(x_t, t) = x_1-x_0$, yielding

$$\mathcal{L}_{\mathrm{FM}}(\theta) = \mathbb{E}_{t, x_0, x_1} \| v_\theta(x_t, t) - (x_1-x_0) \|^2.$$

PCFM augments this setup by introducing a hard constraint $h(x)=0$, with $h: \mathbb{R}^n \to \mathbb{R}^m$ ($m\leq n$), defining the feasible manifold $\mathcal{C} = \{x: h(x)=0\}$. Flow vectors and iterates are projected onto this manifold or its tangent space to ensure constraint satisfaction at all (critical) steps.

For general $h$, a Gauss–Newton projection is used:

$y_{\text{proj}} = y - J^\top (J J^\top)^{-1} r,$

where $r = h(y)$, $J = \nabla h(y)$. This is an orthogonal projection onto the linearized manifold at $y$; exact for affine $h$.

The effective, constraint-respecting velocity is constructed as

$$v_{\mathrm{eff}}(x, t) = \Pi_{T_x \mathcal{C}}\left( v_\theta(x, t) + \Delta(x, t) \right),$$

where the Gauss–Newton correction

$$\Delta(x, t) = -\nabla h(x)^\top \left( \nabla h(x) \nabla h(x)^\top \right)^{-1} \frac{h(x)}{\delta t}$$

is added to drive deviations from the manifold to zero within a single step of size $\delta t$.

The PCFM loss function thus becomes

$$\mathcal{L}_{\mathrm{PCFM}}(\theta) = \mathbb{E} \Big\| \Pi_{T_{x_t}\mathcal{C}}\big( v_\theta(x_t, t) + \Delta(x_t, t) \big) - (x_1-x_0) \Big\|^2,$$

training $v_\theta$ so that the projected, corrected velocity matches the OT interpolant (Utkarsh et al., 4 Jun 2025).

2. Algorithmic Implementation

A typical PCFM inference loop for constraint satisfaction proceeds as follows: at each ODE or Euler discretization step, (1) propagate along the pretrained flow model, (2) project the endpoint onto the constraint manifold by Gauss–Newton projection, (3) reverse and update, and (4) apply a “relaxed correction” to reduce constraint drift. After all steps, a final Newton-Schur solve enforces $h(x)=0$ to machine precision. The algorithm requires only access to the flow, the constraint function, its Jacobian, and supports post-hoc operation on pretrained models.

RelaxedCorrection involves a few (3–5) gradient or Newton steps on an augmented penalty objective that penalizes constraint violation after the relaxed step, maintaining feasibility during coarse integration and improving overall accuracy with very few steps (Utkarsh et al., 4 Jun 2025).

In inverse-imaging applications, e.g., MRI, projection is implemented by orthogonal projection onto the measurement-visible subspace using the pseudoinverse of the acquisition operator, and consistency with noisy measurements is enforced at every backward step (Luo et al., 19 Dec 2025).

3. Theoretical Guarantees and Optimality

If $h$ is $C^2$ and its Jacobian is full row-rank ($m\leq n$) in a neighborhood of $\mathcal{C}$, and the final projection is run to tight numerical tolerance, then the output sample satisfies $h(u(1)) = 0$ to machine precision (Theorem A.1 in (Utkarsh et al., 4 Jun 2025)). Each step’s Gauss–Newton projection reduces residual constraint violation, ensuring fast convergence of the last Newton step initialized in the neighborhood of $\mathcal{C}$.

For ill-posed inverse problems (e.g., MRI) where the measurement map is not full-rank, PCFM defines a projector $P_s = A_s^+A_s$, and the loss is evaluated only on the observable subspace. The solution minimizes the error in that subspace, and the optimal flow-matching vector field is the minimum-MSE estimator conditioned on the measured data (Luo et al., 19 Dec 2025). An explicit link exists between the continuity equation in measurement space and the PCFM objective, yielding a marginal flow ODE to propagate through data space during inference.

4. Applications and Empirical Evaluation

In scientific generative modeling, PCFM was empirically validated on PDE systems: 1D heat equation, 2D incompressible Navier–Stokes, nonlinear reaction–diffusion, and inviscid Burgers’ with shock formation (Utkarsh et al., 4 Jun 2025). Metrics include pointwise MSE for means (MMSE), standard deviation MSE (SMSE), constraint errors (for initial, boundary, and conservation constraints), and Fréchet Poseidon Distance (FPD) via a pretrained PDE encoder.

Key empirical findings:

• PCFM enforces hard constraints (initial, boundary, mass conservation) to machine precision versus persistent residuals for baselines.
• Achieves order-of-magnitude better MMSE/SMSE and FPD on smooth PDEs relative to unconstrained or softly-constrained methods.
• On nonlinear PDEs with shocks, sharply resolves discontinuities while maintaining exact constraint satisfaction.
• Relaxed penalty correction enables accurate results with a small number of steps (10–20), enabling compute–fidelity tradeoff.
• PCFM requires no retraining or architecture changes; it is applied post-hoc.

PCFM has also been applied to unsupervised parallel MRI reconstruction (UPMRI). Here, it enables the learning of fully sampled MRI priors using only undersampled measurement data. Evaluations on fastMRI (brain) and CMRxRecon (cardiac) datasets yielded state-of-the-art reconstruction performance, with UPMRI surpassing traditional supervised, self-supervised, and unsupervised baselines in PSNR and SSIM at high acceleration factors (e.g., 8× brain SSIM: 0.948 vs DDNM⁺: 0.920). Inference required only 20 NFEs (function evaluations) and can be performed at 500 ms/image on modern GPUs (Luo et al., 19 Dec 2025).

5. Comparison with Related Approaches

PCFM differs fundamentally from approaches based on soft penalty enforcement (e.g., PINN-loss guidance as in DiffusionPDE) or architectural bias. Soft penalties do not guarantee physical or mathematical constraints at the solution. Other methods, such as ECI (exact constraint imposition) or gradient-based adjustment (D-Flow), are limited to linear or weakly nonlinear constraints and do not guarantee exactness or may require retraining. PCFM is general, handling arbitrary nonlinear equality constraints and operating in a fully post-hoc fashion.

In inverse problems such as MRI, PCFM’s unsupervised instantiation aligns with the continuity equation and GSURE (generalized Stein’s unbiased risk estimator), allowing principled learning from measurement data alone and recovery of the MMSE estimator in the measurement-visible subspace. This is distinct from self-supervised or plug-and-play techniques, which do not exploit the continuous-time flow-matching interpretation and may not guarantee hard data consistency (Luo et al., 19 Dec 2025).

Method	Guarantees Hard Constraint?	Post-hoc?	Nonlinear Constraints Supported?
PCFM	Yes (machine precision)	Yes	Arbitrary
PINN/Soft Loss	No	No	Yes
ECI/D-Flow	Limited	Yes/No	Mostly linear

6. Practical Considerations and Limitations

PCFM requires that the constraint Jacobian has full row-rank for the tangent-space projection and that the constraint function $h$ is $C^2$. For highly nonlinear or ill-conditioned constraints, the Gauss–Newton step may be suboptimal, but convergence is guaranteed locally if initialized close to the manifold.

In high-dimensional inverse settings, calculating the projection or pseudoinverse may require efficient iterative solvers (e.g., conjugate gradient), which are incorporated into the MRI application. The number of correction steps and Newton iterations may be tuned to trade computation for sample quality or constraint tightness.

A plausible implication is that PCFM provides a principled pathway for integrating domain-specific invariances or conservation laws into black-box generative models across scientific and engineering domains, and enables unsupervised or data-consistent learning in inverse problems where ground truth is unavailable.

7. Impact and Future Directions

PCFM bridges classical projection-based numerical methods with modern deep generative modeling, establishing a new standard for constraint enforcement and data consistency in learning-based simulation and imaging. Its ability to operate post-hoc on pretrained models, require no retraining, and enforce arbitrary nonlinear constraints at inference signals potential for adoption in safety-critical and regulation-compliant domains. Extensions to inequality, soft-constraint, or stochastic settings remain open for further research. The demonstrated empirical advance—state-of-the-art uncertainty-aware generative models and unsupervised MRI reconstruction—suggests broad impact in scientific computing, medical imaging, and inverse problems (Utkarsh et al., 4 Jun 2025, Luo et al., 19 Dec 2025).