Research map / Image generation / Super-resolution
StructSR: Refuse Spurious Details in Real-World 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 StructSR, a training-free plug-and-play inference intervention that screens the highest-SSIM intermediate reconstruction against the LR input and uses it to steer diffusion-based Real-ISR models toward higher structural fidelity and fewer spurious details.
Problem
Diffusion-based real-world image super-resolution often produces structural errors and spurious textures due to the model's empirical priors and illusions, hurting fidelity to the LR input.
Approach
During early diffusion inference, select the intermediate clean latent whose decoded image has the highest SSIM with the bicubic-upsampled LR input (Structure-Aware Screening), then use it as a structural embedding to constrain noise prediction (SCE, weighted by S_max) and to blend into clean latents at each later timestep (IDE, with a timestep-dependent weight), all without fine-tuning.
Key results
- Integrated with four diffusion Real-ISR methods, StructSR improves PSNR and SSIM by an average of 5.27% and 9.36% on DIV2K-Val, and 4.13% and 8.64% on RealSR and DRealSR.
- StableSR+StructSR reaches SSIM 0.6117 on DIV2K-Val versus 0.5335 originally, and DiffBIR+StructSR reaches PSNR 25.09 on RealSR versus 23.67.
- In a 30-participant user study on synthetic data, SeeSR+StructSR obtained a 58.3% confusion rate versus 33.3% for SeeSR alone.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add StructSR to diffusion-based Real-ISR methods | StableSR, DiffBIR, PASD, and SeeSR · DIV2K-Val, RealSR, and DRealSR | PSNR and SSIM | Helped |
| Add StructSR to StableSR | StableSR · RealSR | PSNR: 23.41 -> 24.31 | Helped |
| Add StructSR to DiffBIR | DiffBIR · RealSR | PSNR: 23.67 -> 25.09 | Helped |
| Add StructSR to StableSR | StableSR · DRealSR | SSIM: 0.7252 -> 0.7937 | Helped |
| Add StructSR to DiffBIR | DiffBIR · DRealSR | PSNR: 25.40 -> 27.98 | Helped |
| Add StructSR to DiffBIR | DiffBIR · DIV2K-Val, RealSR, and DRealSR | MUSIQ and CLIP-IQA | Hurt |
| Add StructSR to diffusion-based Real-ISR methods | StableSR, DiffBIR, PASD, and SeeSR · DIV2K-Val, RealSR, and DRealSR | LPIPS | Helped |
| Add StructSR to SeeSR | SeeSR · Synthetic user study | confusion rate: 33.3% -> 58.3% | Helped |
| Add StructSR to SeeSR | SeeSR · Real-world user study | best rate: 11.1% -> 16.7% | Helped |
| Add IDE without SCE (ablation) | StableSR + StructSR · RealSR | PSNR: 23.41 -> 24.29 | Helped |
| Add SCE without IDE (ablation) | StableSR + StructSR · RealSR | SSIM: 0.6824 -> 0.7393 | Helped |
| Add IDE to the SCE-only variant (ablation) | StableSR + StructSR · RealSR | PSNR: 24.07 -> 24.31 | Helped |
Surprising findings
- StructSR lowers perceptual and no-reference scores for DiffBIR (MUSIQ and CLIP-IQA decrease), so gains are concentrated in full-reference metrics.
- GAN-based methods (BSRGAN, LDL) achieve better PSNR, SSIM and LPIPS than the original diffusion baselines on the test sets.
Limitations
- Gains on perceptual metrics (LPIPS, MUSIQ, CLIP-IQA) are mixed and sometimes negative, e.g., DiffBIR.
- Hyperparameter T_SAS=0.3T was chosen from experiments on the same test datasets, and the ablation is in the supplement.
- Relies on full-reference SSIM against the bicubic-upsampled LR image, which may penalize legitimate detail generation.
Experimental setup
Datasets: DIV2K-Val RealSR DRealSR
Benchmarks: DIV2K-Val RealSR DRealSR
Metrics: PSNR SSIM LPIPS MUSIQ CLIP-IQA user study confusion/best rate
Baselines: StableSR DiffBIR PASD SeeSR BSRGAN Real-ESRGAN FeMaSR LDL
Builds on: StableSR DDIM SSIM
Closest papers by meaning
- SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.89)
- Taming Diffusion Prior for Image Super-Resolution with Domain Shift SDEs (NeurIPS 2024 · similarity 0.88)
- Unsupervised Diffusion-Based Degradation Modeling for Real-World Super-Resolution (AAAI 2025 · similarity 0.87)
- BUFF: Bayesian Uncertainty Guided Diffusion Probabilistic Model for Single Image Super-Resolution (AAAI 2025 · similarity 0.87)
- TSD-SR: One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolution (CVPR 2025 · similarity 0.87)
- RAP-SR: RestorAtion Prior Enhancement in Diffusion Models for Realistic Image Super-Resolution (AAAI 2025 · 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.