Research map / Image generation / Fast diffusion sampling
DMPlug: A Plug-in Method for Solving Inverse Problems with Diffusion Models
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 DMPlug, which treats the full DDIM reverse process (with 3 steps) as a fixed function from seed to image and optimizes the seed to solve inverse problems, adding ES-WMV early stopping for unknown-noise robustness.
Problem
Solving inverse problems with pretrained diffusion models while achieving both manifold feasibility and measurement feasibility, and remaining robust to unknown noise types and levels.
Approach
Instead of interleaving reverse diffusion steps with projection/guidance steps, view the whole (deterministic DDIM) reverse process as a fixed function R(z) and optimize the seed z of a data-fitting loss; combined with ELTO-based early stopping (ES-WMV) for noise robustness.
Key results
- On nonlinear deblurring (CelebA), DMPlug reaches PSNR 31.61 vs. ReSample 28.52, about +3.1 dB.
- On BID with motion blur (CelebA), DMPlug reaches PSNR 29.61 vs. best competitor 23.60 (BlindDPS), about +6 dB.
- Under unknown noise types and levels, DMPlug with ES-WMV yields about 1 dB gain for super-resolution and 3.5 dB for nonlinear deblurring over SOTA peak performance, with PSNR gaps under 0.5 dB and 0.2 dB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DMPlug instead of ReSample | Nonlinear deblurring · CelebA, FFHQ, and LSUN | PSNR: +3.09; +5.79; +4.71 | Helped |
| Use DMPlug instead of BlindDPS | BID · CelebA and FFHQ, motion and Gaussian blur | PSNR: +6.01; +3.69; +6.22; +3.92 | Helped |
| Use DMPlug instead of BlindDPS | BID with turbulence · CelebA and FFHQ | PSNR: +2.89; +3.51 | Helped |
| Use DMPlug instead of the strongest baseline | Super-resolution and inpainting · CelebA and FFHQ | PSNR: +2.36; +2.22; +1.79; +2.06 | Helped |
| Use DMPlug instead of DPS | Super-resolution with unknown noise · CelebA, Gaussian, impulse, shot, and speckle noise | PSNR: 0.98; 1.12; 0.87; 1.12 | Helped |
| Use DMPlug instead of DPS | Nonlinear deblurring with unknown noise · CelebA, Gaussian, impulse, shot, and speckle noise | PSNR: 3.61; 3.48; 3.39; 4.23 | Helped |
| Add pretrained DMs for blur kernels and tilt maps (ablation) | DMPlug for BID with turbulence · BID with turbulence | reconstruction quality | Helped |
| Use an LDM prior (ablation) | DMPlug · Super-resolution | reconstruction quality | Helped |
| Replace ADAM with L-BFGS (ablation) | DMPlug · Super-resolution | PSNR: 31.51; 31.89 | No clear effect |
| Increase the number of DDIM reverse steps beyond 3 (ablation) | DMPlug · All evaluated inverse problems | reconstruction quality | Hurt |
| Integrate ES-WMV early stopping | DMPlug · Super-resolution and nonlinear deblurring with unknown noise | PSNR gap: smaller than $0.5$dB; smaller than $0.2$dB | Helped |
Surprising findings
- Only 3 DDIM reverse steps suffice for DMPlug to beat SOTA; more steps slightly degrade performance.
- DMPlug exhibits an early-learning-then-overfitting phenomenon with DM-based methods, which was not previously reported for DM-based IP solvers.
Limitations
- Empirical only; no theoretical understanding of why the seed-optimization approach works.
- Gap between image generation and regression: few reverse steps suffice for IPs but not for photorealistic generation.
- Still slower than some baselines (e.g., DiffPIR, DDRM) and uses ES-WMV with dataset-tuned patience and window sizes.
Experimental setup
Datasets: CelebA FFHQ LSUN-bedroom
Benchmarks: CelebA 256x256 super-resolution FFHQ 256x256 super-resolution LSUN-bedroom super-resolution inpainting 70% random mask nonlinear deblurring blind image deblurring BID with turbulence noise robustness with Gaussian/impulse/shot/speckle noise
Metrics: PSNR SSIM LPIPS PSNR gap
Baselines: DPS ReSample MCG ILVR DMPS DDRM ADMM-PnP BlindDPS BKS-styleGAN BKS-generic Stripformer SelfDeblur
Builds on: DDIM DDPM deep image prior ES-WMV DPS
Closest papers by meaning
- Integrating Intermediate Layer Optimization and Projected Gradient Descent for Solving Inverse Problems with Diffusion Models (ICML 2025 · similarity 0.93)
- SITCOM: Step-wise Triple-Consistent Diffusion Sampling For Inverse Problems (ICML 2025 · similarity 0.90)
- Improving Diffusion-based Inverse Algorithms under Few-Step Constraint via Linear Extrapolation (NeurIPS 2025 · similarity 0.89)
- A Variational Perspective on Solving Inverse Problems with Diffusion Models (ICLR 2024 · similarity 0.89)
- InverseBench: Benchmarking Plug-and-Play Diffusion Priors for Inverse Problems in Physical Sciences (ICLR 2025 · similarity 0.89)
- Self-diffusion for Solving Inverse Problems (NeurIPS 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.