Research map / Image generation / Super-resolution
Uncertainty-Aware GAN for Single Image Super Resolution
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 UGAN, a GAN for perceptual single image super-resolution whose discriminator predicts pixel-level reconstruction uncertainty via uncertainty-aware data augmentation, plus an uncertainty-based loss weighting.
Problem
GAN-based single image super-resolution uses image-level discriminators that output a single global realness score, so they give no fine-grained, pixel-level feedback on hard regions such as textures and edges.
Approach
A U-Net discriminator is trained on images mixing SR and GT pixels, sampled with residual-based probability maps, so it outputs a per-pixel uncertainty/realness map that constrains the generator to focus on high-uncertainty regions; loss weights are learned via homoscedastic uncertainty (Kendall et al.).
Key results
- On BSD100 with SRResNet backbone, UGAN reaches 26.48 PSNR / 0.694 SSIM / 0.089 LPIPS versus 24.38 / 0.620 / 0.204 for the standard GAN.
- With SwinIR backbone on Urban100, UGAN gives 26.39 PSNR / 0.796 SSIM / 0.078 LPIPS versus LDL's 26.231 / 0.7918 / 0.1021.
- On SR quality assessment, the discriminator reaches SROCC 0.787 without training on IQA data, below the trained SISRQA (0.984) but above PSNR, SSIM, FSIM and LPIPS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the RRDBNet GAN baseline with uncertainty-aware adversarial training | RRDBNet · Set14 ×4 | PSNR: 27.228 -> 27.935 | Helped |
| replace the RRDBNet GAN baseline with uncertainty-aware adversarial training | RRDBNet · BSD100 ×4 | LPIPS: 0.1563 -> 0.0859 | Helped |
| replace the RRDBNet GAN baseline with uncertainty-aware adversarial training | RRDBNet · Urban100 ×4 | PSNR: 25.459 -> 25.950 | Helped |
| replace the RRDBNet GAN baseline with uncertainty-aware adversarial training | RRDBNet · DIV2K100 ×4 | SSIM: 0.7951 -> 0.7965 | Helped |
| add feature matching loss to pixel loss (ablation) | SRResNet training · BSD100 ×4 | PSNR: 27.58 -> 26.62 | Hurt |
| add uncertainty-aware adversarial loss to pixel and feature matching losses (ablation) | SRResNet training · Urban100 ×4 | LPIPS: 0.1251 -> 0.0967 | Helped |
| enable uncertainty-aware loss weighting (ablation) | UGAN with SRResNet · Urban100 ×4 | PSNR: 25.18 -> 25.61 | Helped |
| use proportional probability for generator samples and Gaussian probability for discriminator samples instead of no sampling (ablation) | UGAN data augmentation · BSD100 ×4 | PSNR: 25.056 -> 26.479 | Helped |
| use proportional probability for generator samples and Gaussian probability for discriminator samples instead of uniform sampling (ablation) | UGAN data augmentation · BSD100 ×4 | LPIPS: 0.290 -> 0.089 | Helped |
| use patch sampling instead of pixel sampling with uniform probabilities (ablation) | UGAN data augmentation · BSD100 ×4 | PSNR: 26.182 -> 26.707 | Helped |
| use proportional sampling for both generator and discriminator instead of proportional/Gaussian sampling (ablation) | UGAN data augmentation · BSD100 ×4 | LPIPS: 0.261 -> 0.089 | Hurt |
| use Gaussian sampling for both generator and discriminator instead of proportional/Gaussian sampling (ablation) | UGAN data augmentation · BSD100 ×4 | LPIPS: 0.318 -> 0.089 | Hurt |
Surprising findings
- The uncertainty-aware loss weighting does not give the best SR results in the ablation (BSD100 PSNR drops from 26.48 to 26.54 vs. 26.48 without ULW, but LPIPS worsens from 0.089 to 0.158), and the paper frames it as reducing tuning effort rather than improving quality.
Limitations
- The uncertainty-to-quality score for SR-QA is simple averaging and lags trained IQA models (SROCC 0.787 vs 0.984 for SISRQA).
- Uncertainty is not modeled for diffusion-based SR; the authors leave this to future work.
- Results are reported on the authors' TensorFlow implementation with a single GTX 2080Ti, and the ULW component does not clearly improve the best results.
Experimental setup
Datasets: DIV2K Flickr2K
Benchmarks: Set5 Set14 BSD100 Urban100 Manga109 DIV2K100 SIDD SR-QA dataset
Metrics: PSNR SSIM LPIPS SROCC KROCC PLCC RMSE
Baselines: SRGAN ESRGAN RankSRGAN SPSR Beby-GAN LDL SROOE UDL ENLCN HAT
Builds on: SRGAN ESRGAN Kendall multi-task uncertainty loss weighting UDL LDL
Closest papers by meaning
- SeD: Semantic-Aware Discriminator for Image Super-Resolution (CVPR 2024 · similarity 0.88)
- Uncertainty-guided Perturbation for Image Super-Resolution Diffusion Model (CVPR 2025 · similarity 0.88)
- GSDD: Generative Space Dataset Distillation for Image Super-resolution (AAAI 2024 · similarity 0.87)
- Exploring Semantic Feature Discrimination for Perceptual Image Super-Resolution and Opinion-Unaware No-Reference Image Quality Assessment (CVPR 2025 · similarity 0.87)
- Noise-Started One-Step Image Super-Resolution via LR-Conditioned SplitMeanFlow and GAN Refinement (NeurIPS 2026 · similarity 0.87)
- Adversarial Diffusion Compression for Real-World Image Super-Resolution (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.