Research map / Image generation / Image restoration
Improving Diffusion-Based Image Restoration with Error Contraction and Error Correction
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 DiffECC, a diffusion-based image restoration sampler that starts from a CNN restoration output, uses ODE restart sampling for error contraction, and iteratively corrects noise and x0 predictions for error correction, reaching strong SR and deblurring results in about 58 NFEs.
Problem
Diffusion-based solvers for noisy inverse problems in image restoration either give poor quality or need around a thousand sampling iterations to reach high-quality reconstructions.
Approach
Combines a pretrained CNN restoration output (Real-ESRNet) as the diffusion starting point with ODE-based restart sampling to contract errors, and adds an estimation-correction loop on both the noise prediction (Adam-based l1 regularization) and the data-consistency term (DPS-style gradient step).
Key results
- On FFHQ 4x super-resolution, DiffECC reaches 28.47 PSNR and 0.1843 LPIPS with 58 NFEs, versus 24.02 PSNR for DPS with 1000 NFEs.
- On ImageNet 4x super-resolution, DiffECC reaches 23.88 PSNR and FID 43.04 with 58 NFEs, outperforming DiffPIR (22.99 PSNR, FID 55.45) and DDRM (22.36 PSNR).
- Ablation: restart sampling raises ImageNet SR PSNR from 23.97 to 24.68 and lowers FID from 37.59 to 35.52.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DiffECC instead of the strongest listed baseline, DiffPIR | Image restoration on FFHQ · FFHQ 256×256-1k, ×4 SR, σn=0.05 | PSNR: 26.73 -> 28.47 | Helped |
| Use DiffECC instead of the strongest listed baseline, DiffPIR | Image restoration on FFHQ · FFHQ 256×256-1k, Gaussian deblur, σn=0.05 | PSNR: 24.85 -> 26.23 | Helped |
| Use DiffECC instead of the strongest listed baseline, DiffPIR | Image restoration on FFHQ · FFHQ 256×256-1k, motion deblur, σn=0.05 | PSNR: 26.98 -> 27.06 | Helped |
| Use DiffECC instead of DDRM | Image restoration on ImageNet · ImageNet 256×256-1k, ×4 SR, σn=0.05 | PSNR: 22.36 -> 23.88 | Helped |
| Use DiffECC instead of DDRM | Image restoration on ImageNet · ImageNet 256×256-1k, Gaussian deblur, σn=0.05 | PSNR: 22.84 -> 22.35 | Hurt |
| Use DiffECC instead of DiffPIR | Image restoration on ImageNet · ImageNet 256×256-1k, motion deblur, σn=0.05 | PSNR: 13.55 -> 24.04 | Helped |
| Use Real-ESRNet output as the initial prediction rather than x0 initialization at T′=T=50 (ablation) | DiffECC · ImageNet validation, ×4 noisy SR (σn=0.03) | PSNR: 22.38 -> 24.68 | Helped |
| Use Real-ESRNet output as the initial prediction rather than DDIM inversion (ablation) | DiffECC · ImageNet validation, ×4 noisy SR (σn=0.03) | PSNR: 24.55 -> 24.68 | Helped |
| Add restart sampling (ablation) | DiffECC · ImageNet validation, ×4 SR | PSNR: 23.97 -> 24.68 | Helped |
| Add iterative neural-estimation correction to data-consistency correction (ablation) | DiffECC error correction · ImageNet validation, ×4 SR | PSNR: 23.43 -> 24.68 | Helped |
| Use iterative noise correction rather than prediction-derived noise (ablation) | DiffECC error correction · ImageNet validation, ×4 SR | PSNR: 23.91 -> 24.68 | Helped |
Surprising findings
- Using DDIM inversion as initialization gave a much worse FID (96.20) than the simple Real-ESRNet initialization (35.52) in the ImageNet ablation.
Limitations
- Real-world degradations with unknown forward operator are only addressed in the supplementary material, not the main text.
- Relies on a pretrained CNN restorer (Real-ESRNet) whose errors feed into the initialization.
Experimental setup
Datasets: FFHQ ImageNet
Benchmarks: FFHQ 256x256-1k ImageNet 256x256-1k
Metrics: PSNR SSIM LPIPS FID NFE
Baselines: DPS DDRM DDNM DiffPIR
Builds on: DPS DDNM DDIM Xu et al. restart sampling DiffPIR
Closest papers by meaning
- Improving Diffusion Posterior Samplers with Lagged Temporal Corrections for Image Restoration (NeurIPS 2026 · similarity 0.91)
- Residual Denoising Diffusion Models (CVPR 2024 · similarity 0.91)
- A Variational Perspective on Solving Inverse Problems with Diffusion Models (ICLR 2024 · similarity 0.91)
- Resfusion: Denoising Diffusion Probabilistic Models for Image Restoration Based on Prior Residual Noise (NeurIPS 2024 · similarity 0.90)
- Solving Diffusion ODEs with Optimal Boundary Conditions for Better Image Super-Resolution (ICLR 2024 · similarity 0.90)
- A Diffusion Model with State Estimation for Degradation-Blind Inverse Imaging (AAAI 2024 · similarity 0.90)
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.