Research map / Image generation / Image restoration
Uncertainty Quantification via Neural Posterior Principal Components
Elias Nehme, Omer Yair, Tomer Michaeli
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use NPPC instead of a point mass at the conditional mean | posterior uncertainty model · 2D Gaussian mixture denoising | Wasserstein-2 distance: 4.05 -> 0.04 | Improved |
| use NPPC instead of point mass baseline with K=3 components | posterior uncertainty model · 100-dim Gaussian mixture denoising | Wasserstein-2 distance: 180.6 -> 30.3 | Improved |
| use NPPC instead of point mass baseline with K=12 components | posterior uncertainty model · 100-dim Gaussian mixture denoising | Wasserstein-2 distance: 325.7 -> 28.8 | Improved |
| use NPPC instead of DDRM sample PCA | PC-based uncertainty · CelebA-HQ/FFHQ 8x noisy SR | residual error magnitude: 16.00 -> 15.11 | Improved |
| use NPPC instead of DDNM sample PCA | PC-based uncertainty · CelebA-HQ/FFHQ 4x noiseless SR | residual error magnitude: 8.38 -> 8.24 | Improved |
| use NPPC instead of RePaint sample PCA | PC-based uncertainty · CelebA-HQ/FFHQ eyes inpainting | residual error magnitude: 10.3 -> 11.43 | Hurt |
| use NPPC instead of MAT sample PCA | PC-based uncertainty · CelebA-HQ/FFHQ mouth inpainting | residual error magnitude: 11.74 -> 10.27 | Improved |
| use NPPC instead of MAT sampling for speed | PC-based uncertainty · CelebA-HQ/FFHQ posterior-PC comparison | neural function evaluations: 100 -> 1 | Improved |
| add Stopgrad to the Gram-Schmidt projections (ablation) | NPPC joint PC learning · CelebA-HQ colorization | PC consistency (cosine similarity): 0.9 across first 3 PCs, 0.83 overall | Improved |
| train NPPC to predict the mean jointly rather than wrapping a pretrained mean estimator | NPPC · MNIST denoising | model size: 4x | Mixed |
| use more PCs for severely ill-posed tasks (K=10) | NPPC / posterior samplers · CelebA-HQ noisy SR vs inpainting | unexplained variance fraction: 0.8 vs 0.4 | Mixed |
Closest papers by meaning
- Hierarchical Uncertainty Exploration via Feedforward Posterior Trees (NeurIPS 2024 · similarity 0.88)
- Uncertainty Visualization via Low-Dimensional Posterior Projections (CVPR 2024 · similarity 0.88)
- Consistency Posterior Sampling for Diverse Image Synthesis (CVPR 2025 · similarity 0.87)
- BUFF: Bayesian Uncertainty Guided Diffusion Probabilistic Model for Single Image Super-Resolution (AAAI 2025 · similarity 0.87)
- On the Posterior Distribution in Denoising: Application to Uncertainty Quantification (ICLR 2024 · similarity 0.87)
- Uncertainty Estimation for Pretrained Medical Image Registration Models via Transformation Equivariance (NeurIPS 2026 · 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.
