Research map / Image generation / Super-resolution
Noise-Free Optimization in Early Training Steps for 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.
Decomposes HR targets into an optimal centroid and inherent noise, and proposes ECO, which trains SR networks toward a pretrained network's empirical centroid (noise-free target) early in training and mixes in the original HR targets via mixup later.
Problem
Vanilla pixel-wise L1 training of single image super-resolution treats each HR image as the sole ground truth, ignoring the ill-posed nature of SR; the per-sample inherent noise makes early training unstable.
Approach
Replace the per-sample HR target with an empirical centroid estimated by a pretrained SR network (applied consistently to both LR input and target to avoid spatial inconsistency as in KD), then use a mixup schedule (alpha from 0 to 1) to reintroduce the HR target and its high-frequency supervision later in training.
Key results
- ECO improves EDSR at x2 on Set5 from 38.18 dB to 38.29 dB and on Urban100 from 32.83 to 33.07 dB under the reduced-iteration setting.
- ECO with only 20% of training iterations (x8 SR, EDSR-baseline) reaches comparable or better PSNR than vanilla training on Set5, Set14, Urban100, and Manga109 (e.g., Set5 27.00 vs 26.88 dB).
- Gains are also reported for SwinIR-small at x2 (Urban100 32.78 vs 32.74 dB) and for real-world x2 SR with color Gaussian noise.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ECO training with mixup instead of vanilla training | EDSR · Set5 ×2 | PSNR (dB): 38.18 -> 38.29 | Helped |
| Add ECO training with mixup instead of vanilla training | EDSR · Set14 ×2 | PSNR (dB): 33.82 -> 34.07 | Helped |
| Add ECO training with mixup instead of vanilla training | EDSR · BSD100 ×2 | PSNR (dB): 32.33 -> 32.37 | Helped |
| Add ECO training with mixup instead of vanilla training | EDSR · Urban100 ×2 | PSNR (dB): 32.83 -> 33.07 | Helped |
| Add ECO training with mixup instead of vanilla training | EDSR · Manga109 ×2 | PSNR (dB): 39.05 -> 39.26 | Helped |
| Add ECO training with mixup instead of vanilla training | EDSR · Set5 ×3 | PSNR (dB): 34.70 -> 34.80 | Helped |
| Add ECO training with mixup instead of vanilla training | EDSR · Set14 ×3 | PSNR (dB): 30.58 -> 30.64 | Helped |
| Add ECO training with mixup instead of vanilla training | EDSR · BSD100 ×3 | PSNR (dB): 29.26 -> 29.32 | Helped |
| Add ECO training with mixup instead of vanilla training | EDSR · Urban100 ×3 | PSNR (dB): 28.75 -> 28.95 | Helped |
| Add ECO training with mixup instead of vanilla training | EDSR · Manga109 ×3 | PSNR (dB): 34.17 -> 34.36 | Helped |
| Add ECO training with mixup instead of vanilla training | EDSR · Set5 ×4 | PSNR (dB): 32.50 -> 32.59 | Helped |
| Add ECO training with mixup instead of vanilla training | EDSR · Set14 ×4 | PSNR (dB): 28.81 -> 28.90 | Helped |
Surprising findings
- Standalone noise-free training without mixup underperforms vanilla training in later steps, due to neglected estimation error of the pretrained teacher.
- Vanilla knowledge distillation converges slightly faster early but ends worse than vanilla training because of spatial inconsistency between input and target.
Limitations
- Cannot disentangle inherent noise from estimation error, so inherent noise is reintroduced in later training via mixup.
- Standalone noise-free objective hurts final performance because pretrained teacher estimation errors are not optimized.
- Performance gains are modest (often around 0.01-0.3 dB).
Experimental setup
Datasets: DIV2K
Benchmarks: Set5 Set14 BSD100 Urban100 Manga109
Metrics: PSNR SSIM
Baselines: EDSR vanilla training RCAN vanilla training SwinIR vanilla training knowledge distillation
Builds on: EDSR mixup Knowledge Distillation SwinIR
Closest papers by meaning
- Arbitrary-steps Image Super-resolution via Diffusion Inversion (CVPR 2025 · similarity 0.89)
- SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.88)
- ResDiff: Combining CNN and Diffusion Model for Image Super-resolution (AAAI 2024 · similarity 0.88)
- Noise-Started One-Step Image Super-Resolution via LR-Conditioned SplitMeanFlow and GAN Refinement (NeurIPS 2026 · similarity 0.88)
- Towards Realistic Data Generation for Real-World Super-Resolution (ICLR 2025 · similarity 0.87)
- Arbitrary-Scale Image Generation and Upsampling using Latent Diffusion Model and Implicit Neural Decoder (CVPR 2024 · 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.