DC-PnPDP restores ADMM memory, then homogenizes dual-shifted denoiser inputs for pretrained diffusion priors.
ShanghaiTech University, Shanghai Jiao Tong University
Classical ADMM memory is useful for reconstruction, but raw dual residuals violate the AWGN assumption of diffusion denoisers. DC-PnPDP couples the dual variable with a spectral correction layer.
The method separates two roles: restore the feedback that removes bias, then reshape the denoiser input so pretrained diffusion priors remain reliable.
Memoryless PnP diffusion can leave a persistent data-consistency residual. Adding the ADMM dual variable fixes the bias, but the dual-shifted input becomes structured and spectrally colored.
DC-PnPDP keeps the dual feedback for optimization and applies Spectral Homogenization before denoising, avoiding the conflict between data fidelity and prior compatibility.
LACT-90 PSNR gain over the strongest baseline
SVCT-20 PSNR gain over the strongest baseline
fewer denoiser evaluations at matched or better quality
HQS or proximal-gradient PnP diffusion solvers update from instantaneous gradients and discard historical constraint violations.
The dual state accumulates consensus mismatch, acting as integral feedback that drives the data-consistency and prior variables toward agreement.
SH reshapes structured, spectrally colored residuals so the diffusion denoiser receives a statistically compatible pseudo-AWGN input.
DCPnP supplies the optimization memory. SH repairs the statistical interface seen by the denoiser.
DC-PnP keeps the scaled dual variable \( \mathbf{u} \) active, so the solver accumulates consensus error between \( \mathbf{x} \) and \( \mathbf{z} \) instead of relying only on instantaneous gradients.
The dual variable acts like integral feedback: it pushes the fixed point toward consensus \( \mathbf{x}^{*}=\mathbf{z}^{*} \), removing steady-state bias.
The dual-shifted input \( \mathbf{v}=\mathbf{x}+\mathbf{u} \) carries structured residuals from CT streaks or MRI aliasing. SH converts this interface into a pseudo-AWGN input before denoising.
SH keeps the denoiser output closer to the on-manifold reference.
If you find this work interesting or helpful, please cite it.
@inproceedings{du2026dcpnpdp,
title = {Plug-and-Play Diffusion Meets ADMM: Dual-Variable Coupling for Robust Medical Image Reconstruction},
author = {Du, Chenhe and Tian, Xuanyu and Wu, Qing and Liu, Muyu and Yu, Jingyi and Wei, Hongjiang and Zhang, Yuyao},
journal = {Forty-third International Conference on Machine Learning},
year = {2026},
url = {https://openreview.net/forum?id=jEBkuuETjr},
}