Plug-and-play Diffusion Models for Image Compressive Sensing with Data Consistency Projection

Xiaodong Wang, Ping Wang, Zhangyuan Li, Xin YuanarXiv 2025 · DOI · arXiv

Introduction

In the field of computational imaging, researchers frequently encounter "inverse problems," where the goal is to reconstruct a high-quality image ($x$) from corrupted, noisy, or highly compressed measurements ($y$). This challenge is central to technologies like single-pixel imaging, MRI, and remote sensing. Traditionally, two dominant paradigms have emerged to solve these problems: Diffusion Models (DMs) and Plug-and-Play (PnP) methods.

Diffusion Models are generative powerhouses, capable of synthesizing highly realistic images from pure noise. However, when applied to inverse problems, they can sometimes "hallucinate" details that are visually pleasing but physically inaccurate relative to the original measurements. Conversely, PnP methods utilize traditional optimization frameworks to ensure strict data consistency—matching the reconstruction to the physical measurements—but they often rely on simpler image priors that lack the sophisticated representational power of modern deep learning.

This paper introduces a unified framework that merges these two worlds. By decoupling the diffusion process and inserting a principled optimization step, the authors achieve state-of-the-art reconstruction without the need for task-specific retraining.

The Proposed Framework

The core innovation lies in rethinking the Denoising Diffusion Implicit Model (DDIM) sampling process. Rather than treating data consistency as a heuristic "nudge," the authors break down each step of the diffusion "reverse" process into three interpretable stages:

  1. Denoising: A pre-trained generative denoiser estimates a clean image ($x_{0|t}$) from the current noisy state ($x_t$).
  2. Data-Consistency (The Innovation): The estimated clean image is passed through a Hybrid PnP module. This module combines two classic optimization strategies: Generalized Alternating Projection (GAP), which acts as a "hard" constraint forcing the image to match measurements exactly, and Half-Quadratic Splitting (HQS), a "soft" constraint that balances data fidelity with natural image priors.
  3. Sampling: The corrected image ($x'_{0|t}$) is then used to sample the next state in the diffusion chain ($x_{t-1}$).
Comparison of the standard Plug-and-play workflow (a) and the proposed Diffusion model framework (b).

Key Insights

Decoupling and Stabilization

By viewing DDIM as a sequence of "denoise-then-correct" steps, the authors found a natural "plug-in" point for PnP optimization. A significant finding was the "Color Shift" phenomenon: using hard constraints (GAP) alone in diffusion often leads to desaturated or distorted colors. The hybrid approach (GAP + HQS) acts as a stabilizer, ensuring the final output maintains both high-frequency detail and accurate color reproduction.

Computational Efficiency

To handle high-resolution images without the "memory explosion" typical of large sensing matrices, the framework utilizes a 2D compressive sensing model. By employing Kronecker products, the sensing process becomes separable, making the math computationally feasible for practical applications.

Experimental Results

The method was evaluated on the McMaster dataset using pre-trained models from ImageNet. The researchers focused on Single-Pixel Imaging (SPI) across various compression ratios (CR).

At a 5% compression ratio, the hybrid method achieved a PSNR of 24.76 dB, notably outperforming standalone DDIM+GAP (24.09 dB) and DDIM+HQS (24.64 dB). Qualitatively, the hybrid approach produces sharper edges and superior visual fidelity, even at extreme 1% compression levels where physical information is scarce.

Visual comparison of reconstruction results. The hybrid DDIM + GAP + HQS method (right) avoids the color desaturation seen in the GAP-only version and provides sharper details than HQS-only.

Conclusion and Future Work

While the framework sets a new benchmark for zero-shot image reconstruction, the authors identified areas for further refinement. The hybrid method's performance in the very early stages of diffusion is currently a bottleneck, and the fusion weight ($\delta_t$) between hard and soft constraints is currently manually scheduled. Future work will focus on automating parameter tuning and extending this framework to other modalities such as MRI and CT scanning.

Summary updated

BibTeX
@misc{wang2025plug,
  author = {Xiaodong Wang and Ping Wang and Zhangyuan Li and Xin Yuan},
  title = {{Plug-and-play Diffusion Models for Image Compressive Sensing with Data Consistency Projection}},
  year = {2025},
  eprint = {2509.09365},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.09365},
}
Figure