Inverse Flow Matching Problem

• Inverse Flow Matching is a framework that repurposes pretrained flow models as Bayesian priors to recover latent signals from noisy or incomplete measurements.
• The method decomposes the global MAP objective into sequence-wise local updates along the flow trajectory, enabling efficient gradient-based inversion without costly ODE backpropagation.
• Applications include super-resolution, deblurring, inpainting, and MRI compressed sensing, with empirical results demonstrating state-of-the-art performance on high-dimensional problems.

The inverse flow matching problem comprises the application and adaptation of flow matching generative frameworks—originally developed for unconditional synthesis—to the recovery of latent signals, images, or parameters from noisy, indirect, or incomplete measurements. In this context, a pretrained flow matching (FM) model, which transports a tractable base distribution (such as a multidimensional standard normal) to a complex data distribution through the velocity field of a learned ordinary differential equation (ODE), is repurposed as a Bayesian prior within an inverse problem framework. The objective is to solve for posterior or maximum a posteriori (MAP) estimates given an observation model, typically via efficient, differentiable, and theoretically justified algorithms that circumvent the computational bottlenecks of directly optimizing through the flow ODE.

1. Formulation of Inverse Flow Matching

The canonical inverse problem is represented as the recovery of $x_0 \in \mathbb{R}^n$ from observations $y = A x_0 + \epsilon$, with $A \in \mathbb{R}^{m \times n}$ a (potentially ill-posed) linear operator and $\epsilon \sim \mathcal{N}(0, \sigma_y^2 I)$ additive Gaussian noise. In a Bayesian framework, the posterior is $p(x_0|y) \propto p(y|x_0) p_{\mathrm{flow}}(x_0)$, where $p_{\mathrm{flow}}$ is the density induced by a pretrained flow matching model.

The objective in inverse flow matching is either MAP estimation or posterior sampling under this prior. However, direct optimization of $-\log p_{\mathrm{flow}}(x_0)$ is intractable due to the need to backpropagate through a high-dimensional ODE. To address this, approaches such as Iterative Corrupted Trajectory Matching (ICTM) (Zhang et al., 2024), FLOWER (Pourya et al., 30 Sep 2025), and MF-Flow variants employ explicit gradient updates, local approximations, or latent-space reformulations, each leveraging the structure of the flow to enable efficient inversion.

2. Algorithmic Techniques: Trajectory Decomposition and Guidance

A recurring principle in modern inverse flow matching is the decomposition of the global MAP or posterior objective into a sequence of local (stepwise) objectives along the flow trajectory. Specifically, the trajectory parameterized by $t \in [0, 1]$ is partitioned into $N$ steps, each with its own auxiliary observation (often a convex interpolation between the actual measurements and latent noise) and intermediate state $x_t$. The global MAP objective admits an approximation as a weighted sum of “local MAPs,” each involving the instantaneous flow log-determinant, data-fit, and prior terms:

$y = A x_0 + \epsilon$0

where $y = A x_0 + \epsilon$1 are step-dependent weights and each $y = A x_0 + \epsilon$2 is evaluated at the $y = A x_0 + \epsilon$3-th state along the discretized path (Zhang et al., 2024).

The gradient with respect to each $y = A x_0 + \epsilon$4 leverages Tweedie’s formula, expressing the score function $y = A x_0 + \epsilon$5 directly in terms of the flow velocity $y = A x_0 + \epsilon$6 and $y = A x_0 + \epsilon$7 itself. This result enables explicit, stepwise gradient updates on $y = A x_0 + \epsilon$8 without ODE backpropagation:

$y = A x_0 + \epsilon$9

Posterior guidance is incorporated by introducing a likelihood-gradient term (typically, the gradient of the data-fit with respect to the observed measurement operator $A \in \mathbb{R}^{m \times n}$0), which modulates each trajectory update to enforce data consistency.

3. Iterative Corrupted Trajectory Matching (ICTM)

ICTM (Zhang et al., 2024) is an archetypal instance of inverse flow matching for linear problems. The procedure is as follows:

1. Initialization: Start with $A \in \mathbb{R}^{m \times n}$1, set $A \in \mathbb{R}^{m \times n}$2.
2. Auxiliary Measurements: Precompute a sequence of “corrupted” measurements $A \in \mathbb{R}^{m \times n}$3 for $A \in \mathbb{R}^{m \times n}$4.
3. Iterative Updates: For each substep,
   • Predict a forward Euler step: $A \in \mathbb{R}^{m \times n}$5.
   • Compute the gradient of the local MAP objective, combining the prior score, measurement term, and trace regularizer.
   • Take $A \in \mathbb{R}^{m \times n}$6 steps of gradient descent on $A \in \mathbb{R}^{m \times n}$7 for each local objective.
   • Advance $A \in \mathbb{R}^{m \times n}$8, update $A \in \mathbb{R}^{m \times n}$9.
4. Output: Return $\epsilon \sim \mathcal{N}(0, \sigma_y^2 I)$0 as the inverse solution.

No ODE backpropagation is required; each update operates via explicit computation of the required score and measurement terms using only forward evaluations of the flow velocity and its Jacobian-trace.

4. Practical Applications and Task-Specific Instantiation

Inverse flow matching algorithms are designed as universal solvers for a variety of canonical inverse tasks, including:

• Super-resolution: $\epsilon \sim \mathcal{N}(0, \sigma_y^2 I)$1 as downsampling.
• Deblurring: $\epsilon \sim \mathcal{N}(0, \sigma_y^2 I)$2 as convolution with a known kernel.
• Inpainting: $\epsilon \sim \mathcal{N}(0, \sigma_y^2 I)$3 as a masking operator.
• Compressed sensing (MRI): $\epsilon \sim \mathcal{N}(0, \sigma_y^2 I)$4 as partial Fourier measurement.

In each case, the core algorithmic skeleton remains unchanged, with $\epsilon \sim \mathcal{N}(0, \sigma_y^2 I)$5 and its adjoint $\epsilon \sim \mathcal{N}(0, \sigma_y^2 I)$6 replaced by the operator(s) specific to the task. Hyperparameters such as step size $\epsilon \sim \mathcal{N}(0, \sigma_y^2 I)$7, data guidance weight $\epsilon \sim \mathcal{N}(0, \sigma_y^2 I)$8, number of steps $\epsilon \sim \mathcal{N}(0, \sigma_y^2 I)$9, and inner optimization steps $p(x_0|y) \propto p(y|x_0) p_{\mathrm{flow}}(x_0)$0 are tuned for each problem family; empirical results suggest typical values of $p(x_0|y) \propto p(y|x_0) p_{\mathrm{flow}}(x_0)$1, $p(x_0|y) \propto p(y|x_0) p_{\mathrm{flow}}(x_0)$2 to $p(x_0|y) \propto p(y|x_0) p_{\mathrm{flow}}(x_0)$3, $p(x_0|y) \propto p(y|x_0) p_{\mathrm{flow}}(x_0)$4, and $p(x_0|y) \propto p(y|x_0) p_{\mathrm{flow}}(x_0)$5–$p(x_0|y) \propto p(y|x_0) p_{\mathrm{flow}}(x_0)$6 for high-dimensional images (Zhang et al., 2024). Performance empirically scales linearly with the signal dimension.

5. Closed-Form Scores and Tweedie-Style Identities

A central technical advance in inverse flow matching is the replacement of intractable score function evaluations with closed-form expressions grounded in Tweedie’s formula. For the straight-line interpolation flow (with $p(x_0|y) \propto p(y|x_0) p_{\mathrm{flow}}(x_0)$7, $p(x_0|y) \propto p(y|x_0) p_{\mathrm{flow}}(x_0)$8), the score simplifies to:

$p(x_0|y) \propto p(y|x_0) p_{\mathrm{flow}}(x_0)$9

and the corresponding denoising identity is

$p_{\mathrm{flow}}$0

with $p_{\mathrm{flow}}$1, allowing explicit computation of mean estimators and guidance gradients directly from the velocity field. This direct parameterization dramatically reduces the computational complexity compared to diffusion-based methods requiring reverse SDEs or ODE-level backpropagation.

6. Complexity, Scalability, and Theoretical Considerations

Each outer iteration of the standard inverse flow matching algorithm involves $p_{\mathrm{flow}}$2 steps, each requiring a forward network evaluation for the flow and its (trace) Jacobian, and $p_{\mathrm{flow}}$3 inner updates for gradient descent on the current state. The overall complexity is $p_{\mathrm{flow}}$4, where $p_{\mathrm{flow}}$5 denotes the cost of a forward flow (e.g., a UNet) plus trace estimation. Crucially, this architecture allows scaling to high-dimensional problems (e.g., $p_{\mathrm{flow}}$6 images), as the number of neural function evaluations $p_{\mathrm{flow}}$7 need not grow with image resolution provided the flow model is sufficiently expressive.

The method is robust to the choice of noise schedule ($p_{\mathrm{flow}}$8), as long as the path from prior to posterior is Lipschitz. The algorithm provides a consistent approximation for the MAP solution under the flow prior, and the use of explicit closed-form updates avoids the instabilities associated with differentiable ODE solvers or adjoint methods.

7. Empirical Results and Broad Impact

The algorithm is empirically validated on a broad range of linear inverse problems, including super-resolution, deblurring, inpainting, and MRI compressed sensing. Across all tasks, the iterative inverse flow matching algorithm achieves state-of-the-art reconstruction quality and competitive speed, outperforming previous flow-matching-based methods on high-resolution imaging tasks (Zhang et al., 2024). The generality of the framework allows for cross-task applicability with minimal adaptation (replacing $p_{\mathrm{flow}}$9 and $-\log p_{\mathrm{flow}}(x_0)$0), making it an attractive solution for practitioners seeking both theoretical soundness and practical efficiency in solving ill-posed inverse problems.

---

References:

• Flow Priors for Linear Inverse Problems via Iterative Corrupted Trajectory Matching (Zhang et al., 2024)