Research map / Probabilistic and scientific modeling / Variational and generative sampling
pcaGAN: Improving Posterior-Sampling cGANs via Principal Component Regularization
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 pcaGAN, a posterior-sampling cGAN whose regularization adds K principal-component (eigenvector and eigenvalue) terms to the rcGAN mean/trace-covariance objective, and shows gains over cGANs, NPPC and diffusion samplers on denoising, MRI and inpainting.
Problem
Fast and accurate posterior sampling for ill-posed imaging inverse problems; existing cGAN regularizers match only the posterior mean and trace-covariance, and diffusion samplers are orders of magnitude slower.
Approach
Extends rcGAN regularization by additionally matching the top-K principal eigenvectors and eigenvalues of the posterior covariance, using lazy SVD-based estimates from generated samples with stop-gradient on the mean and eigenvalue targets.
Key results
- On MNIST denoising, pcaGAN (K=10) reaches CFID 60.16 vs 63.44 for rcGAN, with REM5 3.25 vs 3.41, while NPPC keeps the best rMSE (3.94 vs 4.02 for pcaGAN).
- On fastMRI at R=8, pcaGAN attains CFID^2 11.47 vs 13.20 for rcGAN and FID 28.35 vs 28.43, with 217 ms generation for 4 samples.
- On FFHQ 256x256 inpainting, pcaGAN achieves CFID 7.08 and FID 1.98, beating DPS (7.26, 2.00) while sampling about 3-4 orders of magnitude faster.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add principal-component regularization to rcGAN | rcGAN · MNIST denoising | CFID: 63.44 -> 60.16 | Helped |
| Add principal-component regularization to rcGAN | rcGAN · MRI R=4 | CFID1: 9.71 -> 8.78 | Helped |
| Add principal-component regularization to rcGAN | rcGAN · MRI R=8 | CFID1: 24.04 -> 21.65 | Helped |
| Add principal-component regularization to rcGAN | rcGAN · FFHQ inpainting | CFID: 7.51 -> 7.08 | Helped |
| Increase estimated components from K=5 to K=10 (ablation) | pcaGAN · MNIST denoising | CFID: 61.48 -> 60.16 | Helped |
| Increase estimated components from K=5 to K=10 (ablation) | pcaGAN · MNIST denoising | REM5: 3.31 -> 3.25 | Helped |
| Increase estimated components from K=5 to K=10 (ablation) | pcaGAN · MNIST denoising | rMSE: 4.02 -> 4.02 | No clear effect |
| Use K<d rather than K=d (ablation) | pcaGAN · Synthetic Gaussian data (d=100) | Wasserstein-2 distance | Hurt |
| Increase lazy update period M (ablation) | pcaGAN · Synthetic Gaussian data (d=100, K=d) | Wasserstein-2 distance | Mixed |
| Use the generated eigenvalue in the denominator rather than the numerator (ablation) | pcaGAN eigenvalue regularization · Not specified | Not specified | Helped |
| Use squared error rather than absolute value for eigenvalue regularization (ablation) | pcaGAN eigenvalue regularization · Not specified | Not specified | Helped |
Surprising findings
- On FFHQ inpainting pcaGAN slightly edges DPS (1000 NFEs) on CFID and FID despite being far faster.
- On MNIST, NPPC achieves the best rMSE, which the authors attribute to its dedicated MMSE network.
Limitations
- Generating P_eig = 10K samples during training is memory-intensive for high-dimensional x, limiting MRI to K=1 at batch size 2.
- MRI application is preliminary and needs more tuning/validation before clinical use.
- Best way to use generated samples for uncertainty quantification remains open.
Experimental setup
Datasets: synthetic Gaussian MNIST fastMRI FFHQ
Benchmarks: MNIST denoising fastMRI accelerated MRI (R=4, R=8) FFHQ 256x256 inpainting synthetic Gaussian W2
Metrics: W2 distance rMSE REM5 CFID FID LPIPS DISTS PSNR SSIM APSD
Baselines: rcGAN NPPC pscGAN Adler & Öktem cGAN Langevin approach (Jalal et al.) E2E-VarNet CoModGAN DPS DDNM DDRM
Builds on: rcGAN NPPC Wasserstein GAN with gradient penalty pscGAN CoModGAN
Closest papers by meaning
- A Restoration Network as an Implicit Prior (ICLR 2024 · similarity 0.87)
- Exact Posterior Score Estimation for Solving Linear Inverse Problems (NeurIPS 2026 · similarity 0.87)
- Consistency Posterior Sampling for Diverse Image Synthesis (CVPR 2025 · similarity 0.87)
- Solving Inverse Problems with FLAIR (NeurIPS 2025 · similarity 0.86)
- Ambient Diffusion Posterior Sampling: Solving Inverse Problems with Diffusion Models Trained on Corrupted Data (ICLR 2025 · similarity 0.86)
- Stochastic Deep Restoration Priors for Imaging Inverse Problems (ICML 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.