Research map / Image generation / Image restoration
Traversing Distortion-Perception Tradeoff using a Single Score-Based Generative Model
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 a variance-scaled reverse diffusion process on a single pre-trained score network that traverses the distortion-perception tradeoff for inverse problems, with a proof of optimality for conditional Gaussians.
Problem
Existing denoising/restoration methods sit at fixed points of the distortion-perception (DP) tradeoff and need retraining or redesign to move to other points when goals or measurement noise change.
Approach
Scale the covariance of the reverse diffusion posterior by a factor lambda in [0,1] while keeping the true conditional mean, so lambda=0 gives the MMSE estimate and lambda=1 gives posterior samples; this traces the DP curve with one score network, and is shown optimal (MSE and W2) for conditional Gaussians.
Key results
- On FFHQ Gaussian deblurring, the variance-scaled sampler spans a broader DP tradeoff than PSCGAN, which needs averaging 64 samples to match the MSE of the lambda=0.9 setting.
- For 8x super-resolution, DiffPIR performs close to the lambda=1 case with better FID but larger distortion than the proposed method.
- For conditional Gaussians, the optimal DP tradeoff gives best MSE of 2Tr(Sigma) at perfect perception, converging to the MMSE value Tr(Sigma) as P grows.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| vary the reverse-sampling variance scale λ | Variance-scaled reverse diffusion · Pinwheel | DP tradeoff coverage | Helped |
| use variance-scaled reverse diffusion instead of PSCGAN | Gaussian deblur, σn = 0.3 · FFHQ | MSE | Helped |
| use variance-scaled reverse diffusion instead of PSCGAN | Gaussian deblur, σn = 0.3 · FFHQ | DP tradeoff coverage | Helped |
| test PSCGAN at σn = 0.5 after training at σn = 0.3 | PSCGAN · Gaussian deblur, σn = 0.5 | FID | Hurt |
| use variance-scaled reverse diffusion instead of DiffPIR | Super-resolution ×8 · FFHQ | MSE | Helped |
| use DiffPIR instead of variance-scaled reverse diffusion | Super-resolution ×8 · FFHQ | FID | Helped |
| increase λ from 0 toward 1 | Variance-scaled reverse diffusion · Gaussian deblur | perceptual quality | Helped |
| increase λ from 0 toward 1 | Variance-scaled reverse diffusion · Gaussian deblur | fidelity | Hurt |
| set reverse-sampling variance scale to λ = 0 instead of λ = 1 | Variance-scaled reverse diffusion · Conditional Gaussian distribution | MSE | Helped |
Surprising findings
- At perfect perception (P=0) the optimal MSE is twice the MMSE value, as shown theoretically and in the 2D toy example.
Limitations
- Optimality is proven only for conditional Gaussian distributions; real image data is evaluated empirically.
- Relies on DPS approximation of the conditional score, which has an upper-bounded Jensen-gap error; hyperparameter zeta needs tuning per lambda.
- Sampling uses the full stochastic reverse SDE with many steps; DDIM-style faster traversal is left for future work.
Experimental setup
Datasets: FFHQ 256x256 pinwheel 2D dataset S-curve 2D dataset moon-shape 2D dataset
Benchmarks: FFHQ 256x256 Gaussian deblur FFHQ 256x256 8x super-resolution 2D pinwheel/S-curve/moon datasets
Metrics: MSE PSNR LPIPS FID Wasserstein-2 distance KL divergence
Baselines: PSCGAN DiffPIR
Builds on: denoising diffusion models (VP-SDE) Diffusion posterior sampling (DPS) Score-based variational inference for inverse problems PSCGAN
Closest papers by meaning
- Aligning Generative Denoising with Discriminative Objectives Unleashes Diffusion for Visual Perception (ICLR 2025 · similarity 0.89)
- Looks Too Good To Be True: An Information-Theoretic Analysis of Hallucinations in Generative Restoration Models (NeurIPS 2024 · similarity 0.88)
- DiracDiffusion: Denoising and Incremental Reconstruction with Assured Data-Consistency (ICML 2024 · similarity 0.88)
- Compressed Image Generation with Denoising Diffusion Codebook Models (ICML 2025 · similarity 0.87)
- Diff-2-in-1: Bridging Generation and Dense Perception with Diffusion Models (ICLR 2025 · similarity 0.87)
- Rethinking Reconstruction and Denoising in the Dark: New Perspective, General Architecture and Beyond (CVPR 2025 · similarity 0.86)
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.