Research map / Image generation / Fast diffusion sampling

DGSolver: Diffusion Generalist Solver with Universal Posterior Sampling for Image Restoration

NeurIPS 2025 · Image generation › Fast diffusion sampling · approach: Fast diffusion ODE solvers · 1 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 DGSolver, a training-free high-order ODE solver with queue-based acceleration and universal posterior sampling for diffusion generalist models, evaluated on universal image restoration.

Problem

Diffusion-based universal image restoration suffers from cumulative discretization errors under few sampling steps, and a trade-off between degradation commonality (unified representation) and restoration quality that existing methods fix with costly compensation mechanisms.

Approach

Derives exact semi-linear ODE solutions for diffusion generalist models (unifying degradations via a linear input term), builds k-th order solvers with a queue-based cache of intermediate terms, and adds universal posterior sampling that corrects the predicted noise using a measurement-consistency gradient, without retraining.

Key results

• Ours-L reaches average 31.36 dB PSNR / 0.920 SSIM across five tasks, versus 29.93 dB / 0.907 for DiffUIR (the strongest universal baseline in Table 1).
• Ablation: universal posterior sampling adds about 0.3 dB average PSNR at k=1 (30.70 to 31.01) and 0.16 dB at k=2 (31.20 to 31.36).
• Ours-T (0.45M params, 5.74 GFLOPs) averages 25.95 dB, while Ours-B (3.65M, 23.97 GFLOPs) reaches 29.16 dB, with performance increasing with model size.

Measured effects reported in the paper

Surprising findings

• More sampling steps do not necessarily improve results; 8 steps performed best, and 9 or 10 steps were slightly worse.
• UPS gains shrink as solver order increases (k=3 with UPS equals k=2 with UPS), suggesting the solver order partly subsumes the correction.

Limitations

• Universal posterior sampling relies on a Jensen-gap approximation with a bound that the authors acknowledge is loose in practice, so its accuracy is not formally guaranteed.
• Performance depends on the chosen degradation commonality parameter delta_T, which varies across tasks and requires tuning.
• Some universal baselines were re-implemented on mixed datasets by the authors, so comparisons depend on re-implementation fidelity.

Experimental setup

Datasets: deraining datasets low-light enhancement datasets desnowing datasets (Snow100K) dehazing datasets deblurring datasets remote sensing restoration datasets (Sen2_MTC)

Benchmarks: Deraining Low-light enhancement Desnowing Dehazing Deblurring Remote sensing restoration (denoising, enhancement, deblur, cloud removal, super-resolution) Snow100K-Real (zero-shot qualitative)

Metrics: PSNR SSIM FLOPs Params

Baselines: SwinIR Restormer MAXIM IR-SDE RDDM AirNet Prompt-IR ProRes IDR AutoDIR DA-CLIP DiffUIR

Builds on: DiffUIR diffusion generalist models diffusion posterior sampling DPS (Chung et al.) PSLD

Closest papers by meaning

• A Variational Perspective on Solving Inverse Problems with Diffusion Models (ICLR 2024 · similarity 0.90)
• GenDeg: Diffusion-based Degradation Synthesis for Generalizable All-In-One Image Restoration (CVPR 2025 · similarity 0.90)
• Improving Diffusion Posterior Samplers with Lagged Temporal Corrections for Image Restoration (NeurIPS 2026 · similarity 0.90)
• Improving Diffusion-Based Image Restoration with Error Contraction and Error Correction (AAAI 2024 · similarity 0.89)
• Diffusion State-Guided Projected Gradient for Inverse Problems (ICLR 2025 · similarity 0.89)
• Ambient Diffusion Posterior Sampling: Solving Inverse Problems with Diffusion Models Trained on Corrupted Data (ICLR 2025 · similarity 0.89)