Research map / Image generation / Fast diffusion sampling

Improving Diffusion Posterior Samplers with Lagged Temporal Corrections for Image Restoration

NeurIPS 2026 · Image generation › Fast diffusion sampling · approach: Diffusion posterior sampling · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes LAMP, a lagged multistep correction that reinterprets diffusion posterior sampling as a discretized forced dynamical system and plugs a temporally filtered data-consistent estimate into DiffPIR and DDRM without extra denoiser evaluations.

Problem

Standard diffusion posterior sampling updates use instantaneous data-consistent estimates, causing temporal variability in reverse dynamics and instability for imaging inverse problems.

Approach

View the PS update as a first-order exponential discretization of PF-ODE plus a residual forcing term, then add a second-order temporal correction using the previous data-consistent estimate (lagged, gamma<0), which can be written as a posterior sampler with a convex-combined estimate D̃_t = (1-β)D_t + βD_{t+Δt}.

Key results

• DDRM-LAMP achieves best PSNR/SSIM/LPIPS in most noisy deblurring and super-resolution settings, e.g., CelebA motion deblur PSNR 25.55 vs DDRM 25.52 and SR x4 PSNR 29.12 vs 28.97.
• DiffPIR-LAMP improves DiffPIR on several tasks, e.g., CelebA Gaussian deblur PSNR 25.04 vs 24.84.
• Naive DDRM-1M and DDRM-2M degrade severely (e.g., FFHQ Gaussian deblur PSNR ~11.8 vs DDRM 25.51), showing plain higher-order updates fail.

Surprising findings

• Naive first- and second-order multistep discretizations (DDRM-1M/2M, DiffPIR-1M/2M) substantially degrade reconstruction, so the residual correction term is essential.
• Gains in PSNR and SSIM are small (often under 0.2 dB), and DiffPIR-LAMP sometimes underperforms its backbone.

Limitations

• Relies on the same assumptions as the backbone PS methods and cannot fix sampler failures or model mismatch.
• Gains are modest and depend on backbone behavior; DiffPIR-LAMP improvements are smaller and inconsistent.
• Introduces hyperparameters (lag strength gamma, warm-up length) that may need tuning.

Experimental setup

Datasets: FFHQ ImageNet CelebA

Benchmarks: FFHQ ImageNet CelebA

Metrics: PSNR SSIM LPIPS

Baselines: DPS DiffPIR DDRM DiffPIR-1M DiffPIR-2M DDRM-1M DDRM-2M

Builds on: DiffPIR DDRM DPM-Solver++ DDIM

Closest papers by meaning

• A Variational Perspective on Solving Inverse Problems with Diffusion Models (ICLR 2024 · similarity 0.91)
• Improving Diffusion-Based Image Restoration with Error Contraction and Error Correction (AAAI 2024 · similarity 0.91)
• Analyzing and Guiding Zero-Shot Posterior Sampling in Diffusion Models (NeurIPS 2026 · similarity 0.91)
• Divide-and-Conquer Posterior Sampling for Denoising Diffusion priors (NeurIPS 2024 · similarity 0.91)
• Provably Robust Score-Based Diffusion Posterior Sampling for Plug-and-Play Image Reconstruction (NeurIPS 2024 · similarity 0.90)
• Ambient Diffusion Posterior Sampling: Solving Inverse Problems with Diffusion Models Trained on Corrupted Data (ICLR 2025 · similarity 0.90)