How Do Diffusion Models Reduce Scanning Redundancy in Ptychography? An Analysis of the PCPS Method
PCPS progressively couples diffusion priors with posterior sampling to achieve efficient and high-accuracy reconstruction in low-overlap ptychography.
Ptychography relies on redundant information from overlapping diffraction measurements for robust reconstruction, but dense scanning leads to low acquisition efficiency. Reducing scan overlap improves efficiency but weakens physical constraints, making reconstruction more ill-posed. The PCPS method proposed in the paper uses diffusion model priors to compensate for the loss of redundancy and dynamically adjusts the balance between exploration and constraint during sampling: broad exploration early, stronger measurement consistency later. Experiments on real X-ray ptychography data show that PCPS outperforms traditional iterative methods and existing diffusion posterior samplers under various overlap conditions, especially in low-overlap scenarios, offering a new path for efficient ptychography.
Ptychography is a computational imaging technique that reconstructs sample images from overlapping diffraction measurements. Its core lies in using the redundant information provided by overlapping regions to achieve robust reconstruction, but dense scanning also causes a surge in acquisition time and data volume. Reducing scan overlap can improve efficiency, but it weakens physical constraints and poses challenges for reconstruction. A recent paper proposed the Progressive Coupling Method for Diffusion Posterior Sampling (PCPS), which attempts to use the prior knowledge of diffusion models to compensate for the loss of measurement redundancy, and achieves better results than traditional iterative methods and existing diffusion posterior samplers.
The Efficiency Dilemma of Ptychography: The Trade-off Between Scanning Redundancy and Reconstruction Quality
Ptychography relies on overlapping diffraction measurements to obtain redundant information, which is the foundation for high-fidelity reconstruction. However, more scan overlap means longer acquisition time and larger data volume. Especially for scenarios such as X-ray ptychography where experimental resources are precious, efficiency becomes a bottleneck. Reducing scan overlap can significantly improve measurement efficiency, but physical constraints are weakened accordingly, and the reconstruction problem becomes more ill-posed. Traditional iterative methods often struggle to recover clear sample structures under low-overlap conditions, so additional prior information is needed to compensate for the missing redundancy.
Ptychographic reconstruction is essentially a phase retrieval problem: the detector records the intensity of the diffraction pattern, while the phase information is lost and must be recovered by computation. The role of scan overlap is to allow adjacent illuminated regions to share information, forming an overdetermined system of equations and thus improving the stability of inversion. When the overlap ratio decreases, the number of independent measurements for each sample region decreases, the system of equations tends toward underdetermined, and the effects of noise and model errors are amplified. This is similar to the trade-off between the number of measurements and sparsity in compressed sensing, but ptychography usually does not explicitly assume that the sample is sparse, so it relies more on redundant measurements to constrain the solution space. Understanding this relationship helps to grasp why low-overlap reconstruction requires strong priors.
Diffusion Models as Priors: Opportunities and Challenges
By learning the data distribution, diffusion models can provide powerful prior information and are expected to guide reconstruction toward a reasonable solution when measurement constraints are insufficient. However, how to effectively use this prior in the posterior sampling process, especially under weak measurement constraints, remains an open problem. Existing diffusion posterior samplers often find it difficult to balance exploring the solution space and maintaining measurement consistency: early on, broad exploration is needed to cover uncertainty, while later, stronger constraints are needed to converge to an accurate solution. The PCPS method proposed in the paper is designed precisely to address this contradiction.
Diffusion posterior sampling is a class of methods that use a pretrained diffusion model as a prior and combine it with the measurement operator at inference time. The basic idea is to use Bayes' formula to decompose the posterior score into a prior score (provided by the diffusion model) and a likelihood score (provided by the measurement consistency constraint). However, the likelihood score is usually difficult to compute exactly, and existing methods mostly use approximations, for example, replacing it with the gradient of the measurement residual. Under weak measurement constraints, such approximation errors are amplified, causing the sampling process either to deviate from the data manifold or to overfit noise. In addition, the prior of the diffusion model is trained on natural images, and there is a domain gap between it and the complex-valued sample distribution in ptychography, which is also a practical challenge that posterior sampling must face.
The PCPS Method: A Progressive Coupling Strategy for Posterior Sampling
The core idea of PCPS is progressive coupling: in the early stage of reconstruction, when uncertainty is high, broader exploration of the solution space is allowed to avoid falling prematurely into local optima; as the reconstruction gradually becomes more refined, the consistency of diffusion guidance is progressively enhanced so that the result moves closer to the measurement data. This dynamic adjustment strategy differs from traditional posterior sampling with fixed weights and can better adapt to the reconstruction process under weak measurement constraints. The paper conducted experiments on real X-ray ptychography data, covering different scan overlap settings, and verified the effectiveness of PCPS.
The idea of progressive coupling is related to annealing and curriculum learning: during optimization or sampling, the strength of constraints is dynamically adjusted so that the system first explores the global structure and then refines local details. In diffusion posterior sampling, this is usually implemented through a guidance coefficient that varies with the sampling step—early coefficients are small, allowing the prior to dominate and enabling broad exploration; later coefficients increase, making measurement consistency dominant. The "coupling" in PCPS also emphasizes that the interaction between the prior and the likelihood is not a simple weighted sum, but a gradual coordination of their contributions along the sampling trajectory. This strategy is conceptually similar to temperature scheduling in Langevin dynamics, but it targets the balance between prior and likelihood in posterior sampling.
Experimental Validation: Significant Improvement Under Low Overlap
Experimental results show that PCPS can consistently improve reconstruction accuracy under multiple scan overlap conditions, outperforming traditional iterative methods and existing diffusion posterior samplers, with especially clear advantages in low-overlap (i.e., weak measurement constraint) scenarios. This verifies that diffusion priors can effectively reduce the measurement redundancy required for ptychographic reconstruction, providing a new path for improving imaging efficiency.
Significance and Outlook: Advancing the Development of Efficient Ptychography
This study demonstrates the potential of diffusion models in computational imaging. It not only provides a practical method for low-redundancy ptychographic reconstruction, but also offers ideas for other imaging techniques limited by measurement efficiency. In the future, how to further optimize the efficiency of using diffusion priors, reduce computational cost, and extend the method to more modalities is worth attention.
One-line Summary
PCPS progressively couples diffusion priors with posterior sampling to achieve efficient and high-accuracy reconstruction in low-overlap ptychography.
Paragraph Summary
Ptychography relies on the redundant information provided by overlapping diffraction measurements for robust reconstruction, but dense scanning leads to low acquisition efficiency. Although reducing scan overlap can improve efficiency, it weakens physical constraints and makes the reconstruction problem more ill-posed. The Progressive Coupling Method for Diffusion Posterior Sampling (PCPS) proposed in the paper uses the prior knowledge of diffusion models to compensate for the loss of redundancy and dynamically adjusts the balance between exploration and constraint during sampling: broad exploration of the solution space early on, and enhanced measurement consistency later. Experiments on real X-ray ptychography data show that PCPS outperforms traditional iterative methods and existing diffusion posterior samplers under different overlap conditions, with especially significant advantages in low-overlap scenarios, providing a new path for efficient ptychography.
Related articles
world2vec: Building Mechanistic World Models with Latent Dynamics Learning
world2vec proposes a reward-free method for learning mechanistic world models via latent dynamics, faithfully reproducing systems over long horizons with at least an order of magnitude fewer parameters than DreamerV3, TD-MPC2, and LeWM, and achieving accurate system identification on 71 attractor benchmarks.
A New KL-DRO Paradigm: Adaptive Scenario Generation Around a Learned Density Model
KL-DRO research proposes centering the KL ball on a learned density model instead of the empirical distribution, with adaptive scenario generation guiding sampling toward the worst case. Experiments show lower cost under distribution shift when records are scarce and the model can extrapolate, with comparable in-distribution cost.
ODE Samplers for Discrete Diffusion Models: Exponential Residual Flows Accelerate Generation
Discrete diffusion models lack deterministic ODE sampling? This article introduces Exponential Residual Flows (ERFs), which introduce continuous residual variables per token to construct piecewise ODEs for deterministic sampling and support consistency distillation acceleration. Superior to prior methods on code and text generation tasks.