Technical Summary: FlowLPS

Problem Statement

The paper addresses the challenge of solving imaging inverse problems (recovering a signal $x_0$ from noisy measurements $y = A(x_0) + n$) using pretrained latent flow models. While deep generative models serve as powerful priors, training-free solvers face a practical trade-off under a finite reverse-time budget (limited computational steps).

Existing approaches generally fall into two regimes with distinct limitations in nonlinear latent spaces (e.g., those governed by VAE decoders):

1. Refinement-dominant solvers: These methods (e.g., FLAIR, Flower) aggressively enforce measurement consistency via local optimization. While they rapidly reduce measurement error, their results are often sensitive to initialization and can perturb the prior-supported trajectory, leading to artifacts or degraded perceptual realism.
2. Sampling-dominant solvers: These methods (e.g., DAPS) rely on stochastic Langevin dynamics to explore the posterior. While they preserve prior fidelity and stochasticity, they often require excessive iterations to concentrate into sharp, measurement-consistent estimates, making them inefficient under fixed budgets.

The core problem is the lack of a method that effectively balances stochastic posterior exploration with deterministic measurement-driven refinement within a limited number of inference steps.

Methodology: FlowLPS

The authors propose FlowLPS (Langevin-Proximal Sampling), a training-free hybrid solver designed to bridge the gap between the two operating regimes. The method operates at each reverse-time step $t$ through a three-stage process:

1. Posterior-Oriented Langevin Initialization

Instead of using Langevin dynamics as a long-run posterior sampler, FlowLPS uses a few Langevin updates ($N_L$ steps) to perturb the model-predicted clean estimate ($\hat{z}_{0|t}$).

• Mechanism: Starting from the flow model's prediction, the algorithm performs stochastic gradient updates guided by the conditional posterior gradient $\nabla_{z_0} \log p(z_0|z_t, y)$.
• Goal: To generate a diverse, posterior-oriented initialization that explores the local basin of attraction, rather than relying solely on the deterministic trajectory of the flow model.

2. Local MAP-Style Proximal Refinement

Following the stochastic initialization, FlowLPS applies a local Maximum A Posteriori (MAP) refinement.

• Mechanism: The algorithm solves a proximal optimization problem initialized at the Langevin-updated point. Under Gaussian assumptions for the prior and noise, this is formulated as minimizing a composite objective:
  $$\min_{z_0} \|y - A(z_0)\|^2 + \frac{\sigma_n^2}{s_t^2} \|z_0 - \hat{z}_{0|t}^{(NL)}\|^2$$
  where the first term enforces measurement consistency and the second term acts as a proximal regularizer keeping the solution close to the stochastic initialization.
• Goal: To rapidly concentrate the estimate into a sharp, measurement-consistent reconstruction, leveraging the "good" initialization provided by the Langevin phase.

3. Controlled pCN-Style Re-noising

To stabilize the reverse trajectory after refinement, FlowLPS employs a controlled re-noising step inspired by the preconditioned Crank-Nicolson (pCN) algorithm.

• Mechanism: Instead of drawing a completely new noise vector for the next step, the noise component is refreshed via a linear combination of the previous noise estimate and new Gaussian noise: $\hat{z}_{1|t}^{(new)} = \rho_t \hat{z}_{1|t}^{(old)} + \sqrt{1-\rho_t^2}\epsilon$.
• Goal: To inject necessary stochasticity to prevent trajectory collapse while retaining coherence with the model-predicted noise direction, balancing exploration and stability.

Key Contributions

1. FlowLPS Framework: A training-free latent flow inverse solver that integrates posterior-oriented Langevin initialization with local measurement-consistent proximal refinement.
2. Theoretical Analysis: A local analysis demonstrating the complementary roles of the two phases: Langevin updates provide stochastic exploration to escape poor local minima, while proximal refinement ensures geometric contraction toward the local minimizer within a finite step budget.
3. Stabilization Mechanism: The introduction of controlled pCN-style re-noising to maintain trajectory coherence while refreshing the noise component.
4. Empirical Validation: Comprehensive experiments showing that FlowLPS achieves a superior balance between distortion metrics (PSNR, SSIM) and perceptual quality (FID, LPIPS) compared to state-of-the-art baselines.

Experimental Results

The authors evaluated FlowLPS on the FFHQ and DIV2K datasets across five linear inverse problems: Gaussian deblurring, Motion deblurring, Super-resolution ($\times 12$), Box Inpainting, and Random Inpainting.

• Performance Balance: FlowLPS consistently outperforms or matches refinement-dominant methods (like Flower) in perceptual metrics (FID, LPIPS) while maintaining competitive distortion metrics. Conversely, it significantly outperforms sampling-dominant methods (like DAPS) in both distortion and perceptual metrics under matched computational budgets (40 NFEs).
• Ablation Studies:
  • Langevin Steps ($N_L$): An intermediate number of Langevin steps (e.g., $N_L=5-7$) yields the best results. $N_L=0$ leads to over-smoothed results (refinement-dominant), while excessive $N_L$ leads to noisy, unstable details (sampling-dominant).
  • pCN Schedule: An adaptive mixing schedule ($\rho_t = \sqrt{1-t}$) outperforms fixed schedules, suggesting that stochastic refresh should depend on the remaining reverse-time uncertainty.
• Generalizability: Additional experiments on pixel-space inverse problems and phase retrieval confirm that the Langevin-Proximal principle is not limited to latent-space linear problems.

Significance and Claims

The paper claims that FlowLPS addresses a critical practical gap in inverse problem solving: the trade-off between stochastic exploration and deterministic refinement under finite computational budgets.

• Practical Efficiency: By treating Langevin updates as a mechanism for initialization rather than full posterior sampling, FlowLPS avoids the high cost of extensive stochastic sampling while avoiding the initialization sensitivity of pure optimization.
• Perception-Distortion Trade-off: The method successfully navigates the perception-distortion trade-off, producing reconstructions that are both mathematically consistent with measurements and perceptually realistic, avoiding the "over-smoothed" artifacts of aggressive optimizers and the "noisy" artifacts of insufficiently converged samplers.
• Limitations: The authors modestly note that FlowLPS depends on the pretrained backbone prior; thus, highly out-of-distribution inputs may still produce artifacts or hallucinated details, requiring caution in high-stakes applications like medical or forensic imaging.