Research map / Image generation / Super-resolution
Image Super-Resolution with Guarantees via Conformalized Generative Models
Eduardo Adame, Daniel Csillag, Guilherme Tegoni Goedert
Read the paperCheck your idea against 137,217 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 conformal "confidence masks" for generative image super-resolution that guarantee expected fidelity error (and PSNR) control on any black-box model using unlabelled calibration images.
Problem
Quantifying where generated super-resolution images can be trusted, with interpretable, statistically guaranteed uncertainty for black-box generative models.
Approach
Calibrate a threshold on a smoothed per-pixel model-indecision map (variance computed via a low-pass kernel on second moments) using conformal risk control, so that the expected supremum of a local fidelity error over the trusted binary mask is at most alpha; also proves PSNR lower bounds and bounds under calibration data leakage.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply conformal masking at α=0.100 | SinSR super-resolution · Liu4K test, semantic D_p | PSNR: 26.83 ± 1.06 -> 32.65 ± 1.48 | Improved |
| Apply conformal masking at α=0.100 | SinSR super-resolution · Liu4K test, non-semantic D_p | PSNR: 26.82 ± 1.08 -> 28.64 ± 0.93 | Improved |
| Apply conformal masking at α=0.100 | SinSR super-resolution · Liu4K test, semantic D_p | fidelity error: 0.14 ± 0.01 -> 0.03 ± 0.01 | Improved |
| Apply conformal masking at α=0.100 | SinSR super-resolution · Liu4K test, non-semantic D_p | fidelity error: 0.14 ± 0.01 -> 0.09 ± 0.01 | Improved |
| Use a neighborhood-averaged, low-pass-filtered D_p instead of pointwise D_p (ablation) | Conformal mask · Image super-resolution | conformal mask size | Improved |
| Increase the Gaussian blur radius for D_p (ablation) | Conformal mask · Image super-resolution | trusted-region coverage | Improved |
| Increase the Gaussian blur radius for σ (ablation) | Conformal mask · Image super-resolution | trusted-region coverage | Improved |
| Use pixel-wise variance σ^var without patch-based low-pass filtering (ablation) | Conformal mask · Image super-resolution | conformal mask quality | Hurt |
| Calibrate on DIV2K distribution-shifted data instead of in-distribution Liu4K | Conformal mask · DIV2K validation | PSNR: 13.34 ± 6.91; 13.14 ± 6.55; baseline 11.73 ± 0.39; 11.74 ± 0.39 | Mixed |
Surprising findings
- A published conformal-style uncertainty method (the continuous confidence-mask approach) is shown to lack its claimed guarantee via a counterexample, since its calibration omits the expectation and uses a plain empirical quantile.
- Smoothing the fidelity error D_p with a low-pass filter produces larger masks than the pointwise L1 error.
Limitations
- Guarantees rely on exchangeability between calibration data and test samples; significant distribution shift may void them.
- Requires generating multiple high-resolution samples per image, which increases computation time.
- Model indecision sigma is not estimated jointly with the upscaled image, left for future work.
Experimental setup
Datasets: Liu4K DIV2K
Benchmarks: Liu4K validation (400 images) DIV2K validation (distribution shift)
Metrics: fidelity error PSNR conformal mask size
Baselines: interval-valued conformal super-resolution method (unnamed in text) continuous confidence-mask method (unnamed in text) no uncertainty quantification
Builds on: conformal risk control conformal prediction
Closest papers by meaning
- The Perception-Robustness Tradeoff in Deterministic Image Restoration (ICML 2024 · similarity 0.87)
- Universal Robustness via Median Randomized Smoothing for Real-World Super-Resolution (CVPR 2024 · similarity 0.86)
- Towards Realistic Data Generation for Real-World Super-Resolution (ICLR 2025 · similarity 0.86)
- Consistency Posterior Sampling for Diverse Image Synthesis (CVPR 2025 · similarity 0.86)
- BUFF: Bayesian Uncertainty Guided Diffusion Probabilistic Model for Single Image Super-Resolution (AAAI 2025 · similarity 0.86)
- Robustness-Guided Image Synthesis for Data-Free Quantization (AAAI 2024 · similarity 0.85)
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.
