Research map / Image generation / Image restoration
DiracDiffusion: Denoising and Incremental Reconstruction with Assured Data-Consistency
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 DiracDiffusion, which models the observation as a sample from a stochastic degradation process and learns to reverse it with an incremental reconstruction loss, enforcing data consistency throughout sampling.
Problem
Diffusion-based inverse problem solvers achieve high perceptual quality but poor distortion metrics (PSNR/SSIM) and often fail to maintain data consistency with the measurement.
Approach
Treat the noisy measurement as a sample from the forward process of a stochastic degradation process (degradation severity indexed by t), train a score network that predicts the clean image in the degraded domain, and add an incremental reconstruction loss so each reverse step reverses a small degradation step, giving built-in data consistency and early-stopping trade-offs.
Key results
- Deblurring on FFHQ: Dirac-DO reaches 28.47 PSNR and 0.8054 SSIM, close to SwinIR's 28.53 PSNR, while Dirac-PO gets the best LPIPS (0.2716) and FID (53.36) among diffusion methods.
- Inpainting on FFHQ: Dirac-DO reaches 26.98 PSNR, 0.8435 SSIM and 0.2234 LPIPS, the best in all three of these metrics, and Dirac-PO reaches the best FID (39.43).
- On ImageNet deblurring, Dirac-DO achieves 25.76 PSNR and 0.7085 SSIM, beating all compared methods on distortion metrics.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Dirac-DO instead of Dirac-PO | Dirac on deblurring · FFHQ test split | PSNR: 26.67 -> 28.47 | Helped |
| Use Dirac-DO instead of Dirac-PO | Dirac on inpainting · FFHQ test split | PSNR: 25.41 -> 26.98 | Helped |
| Use Dirac-DO instead of Dirac-PO | Dirac on deblurring · ImageNet test split | PSNR: 24.68 -> 25.76 | Helped |
| Use Dirac-DO instead of Dirac-PO | Dirac on inpainting · ImageNet test split | PSNR: 26.36 -> 28.92 | Helped |
| Increase incremental reconstruction loss parameter Δt (ablation) | Dirac on inpainting · Inpainting | reconstruction quality | Helped |
| Vary incremental reconstruction loss parameter Δt (ablation) | Dirac on deblurring · Deblurring | reconstruction quality | No clear effect |
| Add measurement-guidance term (ablation) | Dirac on deblurring · CelebA-HQ validation set | LPIPS | Helped |
| Change guidance step-size scheduling (ablation) | Dirac on deblurring · CelebA-HQ validation set | LPIPS | No clear effect |
| Early-stop the reverse process (ablation) | Dirac · Inverse problems | PSNR | Mixed |
Surprising findings
- Distortion metrics peak early in the reverse process and then deteriorate while perceptual metrics keep improving, so early stopping trades perception for distortion.
Limitations
- A model must be trained from scratch for each inverse problem, unlike methods that reuse pretrained score networks.
- Theoretical data-consistency guarantee holds only under ideal assumptions (perfect score and incremental reconstruction).
- Evaluated only on two degradation types (Gaussian blur and inpainting) and two datasets at 256x256.
Experimental setup
Datasets: CelebA-HQ ImageNet FFHQ
Benchmarks: CelebA-HQ ImageNet FFHQ
Metrics: PSNR SSIM LPIPS FID
Baselines: DPS DDRM SwinIR PnP-ADMM ADMM-TV
Builds on: score-based generative models DPS Cold Diffusion Soft Diffusion NCSN++
Closest papers by meaning
- A Variational Perspective on Solving Inverse Problems with Diffusion Models (ICLR 2024 · similarity 0.90)
- Improving Diffusion Posterior Samplers with Lagged Temporal Corrections for Image Restoration (NeurIPS 2026 · similarity 0.88)
- Solving Video Inverse Problems Using Image Diffusion Models (ICLR 2025 · similarity 0.88)
- A Diffusion Model with State Estimation for Degradation-Blind Inverse Imaging (AAAI 2024 · similarity 0.88)
- Adapt and Diffuse: Sample-adaptive Reconstruction via Latent Diffusion Models (ICML 2024 · similarity 0.88)
- Diffusion State-Guided Projected Gradient for Inverse Problems (ICLR 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.