Research map / Image generation / Fast diffusion sampling
Diffusion Posterior Sampling for General Noisy Inverse Problems
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Approximate the time-dependent measurement likelihood using the likelihood of the Tweedie posterior-mean estimate, without measurement-subspace projections | Diffusion posterior sampling · Noisy linear inverse problems on FFHQ and ImageNet | FID: 39.35 vs 62.15; 50.66 vs 59.57 | Helped |
| Use DPS for nonlinear phase retrieval instead of HIO | Phase retrieval · Phase retrieval, FFHQ | FID: 55.61 vs 96.40 | Helped |
| Use DPS for nonlinear deblurring instead of BKS-styleGAN2 | Non-uniform deblurring · Non-uniform deblurring, FFHQ | FID: 41.86 vs 63.18 | Helped |
| Replace the constant step size 1.0 with constant step size 1/σ² (ablation) | DPS · Gaussian deblurring, FFHQ, Gaussian noise | LPIPS: 0.247 ± 0.045 -> 0.727 ± 0.038 | Hurt |
| Replace the constant step size 1.0 with linear decay initialized at 0.3 (ablation) | DPS · Gaussian deblurring, FFHQ, Gaussian noise | LPIPS: 0.247 ± 0.045 -> 0.287 ± 0.045 | Hurt |
| Replace the constant step size 1.0 with linear decay initialized at 1.0 (ablation) | DPS · Gaussian deblurring, FFHQ, Gaussian noise | LPIPS: 0.247 ± 0.045 -> 0.251 ± 0.044 | No clear effect |
| Replace the constant step size 1.0 with exponential decay initialized at 0.3 (ablation) | DPS · Gaussian deblurring, FFHQ, Gaussian noise | LPIPS: 0.247 ± 0.045 -> 0.421 ± 0.065 | Hurt |
| Replace the constant step size 1.0 with exponential decay initialized at 1.0 (ablation) | DPS · Gaussian deblurring, FFHQ, Gaussian noise | LPIPS: 0.247 ± 0.045 -> 0.442 ± 0.108 | Hurt |
| Add measurement-subspace projection steps to the noisy inverse-problem sampling path (ablation) | DPS-style diffusion sampling · Noisy super-resolution and deblurring | reconstruction quality | Hurt |
| Use Poisson-LS instead of the proposed Poisson-shot likelihood treatment (ablation) | DPS · Poisson linear inverse problems | high-frequency detail preservation | Hurt |
| Use the direct Poisson likelihood (ablation) | DPS · Poisson linear inverse problems | sampling stability | Hurt |
| Use the Poisson-Gaussian likelihood approximation (ablation) | DPS · Poisson linear inverse problems | sampling convergence | Hurt |
Closest papers by meaning
- Improving Diffusion Inverse Problem Solving with Decoupled Noise Annealing (CVPR 2025 · similarity 0.92)
- A Variational Perspective on Solving Inverse Problems with Diffusion Models (ICLR 2024 · similarity 0.92)
- Improving Diffusion-based Inverse Algorithms under Few-Step Constraint via Linear Extrapolation (NeurIPS 2025 · similarity 0.91)
- Unleashing the Denoising Capability of Diffusion Prior for Solving Inverse Problems (NeurIPS 2024 · similarity 0.91)
- Improving Diffusion Models for Inverse Problems Using Optimal Posterior Covariance (ICML 2024 · similarity 0.91)
- Linearly Constrained Diffusion Implicit Models (NeurIPS 2025 · similarity 0.91)
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.