Research map / Image generation / Super-resolution
Training Generative Image Super-Resolution Models by Wavelet-Domain Losses Enables Better Control of Artifacts
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.
Trains a GAN-based ×4 super-resolution generator (ESRGAN/RRDB) with SWT-domain L1 fidelity and an SWT detail-subband discriminator (WGSR) to improve perception-distortion trade-off and control artifacts.
Problem
GAN-based super-resolution generates hallucinations and artifacts while reconstructing high-frequency details, and standard fidelity/perceptual metrics fail to indicate such artifacts; existing frequency-domain losses do not localize detail by scale and orientation.
Approach
Compute the generator's L1 fidelity loss over stationary wavelet transform (SWT) subbands with per-subband weights, and train the discriminator only on high-frequency detail subbands (LH, HL, HH) of the Y channel, applied to an RGB-domain RRDB generator with DISTS perceptual loss.
Key results
- On BSD100, the ablation combining SWT-domain L1, SWT-domain adversarial loss and DISTS (#5) reaches 26.331 dB PSNR vs 24.506 dB for the ESRGAN-style baseline (#0), with PI improving from 2.543 to 2.453.
- WGSR (1-level SWT) reports PSNR 26.689 and NRQM 7.927 on Set14-style table entries, competitive with GAN-SR SOTA; on Urban100 it reports NRQM 7.406, the best among compared methods in the table.
- Choice of wavelet family matters; sym7 was selected as the best trade-off point on BSD100 PSNR-NRQM plane.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace LPIPS with DISTS perceptual loss (ablation) | ESRGAN baseline (#0) · BSD100 | PSNR: 24.506 -> 24.812 | Helped |
| replace LPIPS with DISTS perceptual loss (ablation) | ESRGAN baseline (#0) · BSD100 | PI: 2.543 -> 2.502 | Helped |
| compute fidelity loss in SWT domain (ablation) | WGSR with RGB adversarial loss and DISTS (#1) · BSD100 | PSNR: 24.812 -> 25.622 | Helped |
| compute fidelity loss in SWT domain (ablation) | WGSR with RGB adversarial loss and DISTS (#1) · BSD100 | PI: 2.502 -> 2.501 | No clear effect |
| compute adversarial loss in SWT domain (ablation) | WGSR with RGB fidelity loss and DISTS (#1) · BSD100 | PI: 2.502 -> 2.443 | Helped |
| use SWT-domain fidelity and adversarial losses with LPIPS (ablation) | ESRGAN baseline (#0) · BSD100 | PSNR: 24.506 -> 25.859 | Helped |
| use SWT-domain fidelity and adversarial losses with LPIPS (ablation) | ESRGAN baseline (#0) · BSD100 | PI: 2.543 -> 2.466 | Helped |
| replace LPIPS with DISTS in the SWT-loss model (ablation) | WGSR variant #4 · BSD100 | PSNR: 25.859 -> 26.331 | Helped |
| replace LPIPS with DISTS in the SWT-loss model (ablation) | WGSR variant #4 · BSD100 | PI: 2.466 -> 2.453 | Helped |
| use 2-level instead of 1-level SWT (ablation) | WGSR · Urban100 | PSNR: 29.188 -> 29.857 | Helped |
| use 2-level instead of 1-level SWT (ablation) | WGSR · Set14 | PSNR: 27.395 -> 26.689 | Hurt |
| use 2-level instead of 1-level SWT (ablation) | WGSR · BSD100 | PSNR: 26.471 -> 26.372 | Hurt |
Surprising findings
- Standard PSNR, DISTS and other quantitative scores are not good indicators of the visual artifacts that WGSR suppresses, as shown in the Figure 2 example.
Limitations
- Selecting per-subband loss weights and PD trade-off is difficult and requires empirical search.
- Quantitative metrics do not reflect visual artifact suppression.
- Perceptual and PSNR gains are mixed across benchmarks; some baselines have higher PSNR on certain datasets.
Experimental setup
Datasets: DIV2K DF2K Set5 Set14 BSD100 Urban100
Benchmarks: Set5 Set14 BSD100 Urban100 DIV2K validation
Metrics: PSNR SSIM LPIPS LPIPS-VGG DISTS NIQE NRQM PI LR-PSNR
Baselines: ESRGAN-FS ESRGAN+ SPSR RankSRGAN SRFlow-DA LDL FxSR PDASR DualFormer SROOE
Builds on: ESRGAN RRDB DISTS SWT
Closest papers by meaning
- Auto-Encoded Supervision for Perceptual Image Super-Resolution (CVPR 2025 · similarity 0.88)
- SeD: Semantic-Aware Discriminator for Image Super-Resolution (CVPR 2024 · similarity 0.88)
- Noise-Started One-Step Image Super-Resolution via LR-Conditioned SplitMeanFlow and GAN Refinement (NeurIPS 2026 · 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)
- GSDD: Generative Space Dataset Distillation for Image Super-resolution (AAAI 2024 · similarity 0.87)
- When Latents Forget Pixels: Restoring Fidelity in Diffusion Transformer Super-Resolution (NeurIPS 2026 · similarity 0.87)
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.