Research map / Image generation / Super-resolution
Self-Adaptive Reality-Guided Diffusion for Artifact-Free 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 SARGD, a training-free diffusion super-resolution method that uses a PAL artifact detector to mask artifacts in latent space, refines them with realistic-latent guidance, and self-adapts that guidance via a reality score, halving sampling steps.
Problem
Diffusion-based super-resolution produces artifacts (noise, unauthentic textures) during iterative sampling and over-smoothing when using upscaled LR latents as guidance.
Approach
Combines an artifact detector applied to decoded intermediate latents (binary mask) with replacing masked latent regions by a realistic latent guidance (Reality-Guided Refinement), and a reality-score-gated update of that guidance (Self-Adaptive Guidance), all without training.
Key results
- On Set5 at x2, SARGD reaches 32.64 PSNR versus 28.78 for StableSR, a gain of 3.86 dB.
- SARGD raises Set14 x2 SSIM from 0.716 (StableSR) to 0.842.
- Peak PSNR is reached at about 100 of 200 sampling steps, halving inference time relative to StableSR's 200-step setting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SARGD instead of StableSR | Diffusion-based super-resolution · Set5 ×4 | PSNR: 29.36 -> 32.27 | Helped |
| use SARGD instead of StableSR | Diffusion-based super-resolution · Set14 ×4 | PSNR: 28.04 -> 30.01 | Helped |
| use SARGD instead of StableSR | Diffusion-based super-resolution · B100 ×4 | PSNR: 28.40 -> 30.23 | Helped |
| use SARGD instead of StableSR | Diffusion-based super-resolution · Urban100 ×4 | PSNR: 26.55 -> 27.93 | Helped |
| use SARGD instead of StableSR | Diffusion-based super-resolution · Manga109 ×4 | PSNR: 27.94 -> 30.23 | Helped |
| add Reality-Guided Refinement (RGR) (ablation) | LDM baseline · Set5 ×4 | PSNR: 28.30 -> 30.27 | Helped |
| add Reality-Guided Refinement (RGR) (ablation) | LDM baseline · Set14 ×4 | PSNR: 27.21 -> 27.79 | Helped |
| add Reality-Guided Refinement (RGR) (ablation) | LDM baseline · B100 ×4 | PSNR: 27.67 -> 28.30 | Helped |
| add Reality-Guided Refinement (RGR) (ablation) | LDM baseline · Manga109 ×4 | PSNR: 27.04 -> 27.88 | Helped |
| refine artifacts after 100 steps rather than before 100 steps (ablation) | SARGD RGR · Set14 ×4 | PSNR: 26.94 -> 28.33 | Helped |
| update realistic latent after 100 steps rather than every step (ablation) | SARGD SAG · Set14 ×4 | PSNR: 29.28 -> 30.01 | Helped |
| add artifact detection to latent refinement (ablation) | SARGD · B100 ×3 | PSNR: 30.06 -> 30.84 | Helped |
Surprising findings
- Refining artifacts only in the last 100 denoising steps gives better results than refining at every step, since early refinement causes over-smoothing.
Limitations
- Relies on a pretrained artifact detector (PAL) whose errors propagate into the mask.
- Gains on some LPIPS/DISTS entries are mixed, with some metrics worse than baselines at certain scales.
- Evaluation limited to bicubic-downsampled benchmark datasets.
Experimental setup
Benchmarks: Set5 Set14 B100 Urban100 Manga109
Metrics: PSNR SSIM LPIPS DISTS
Baselines: StableSR LDM Bicubic
Builds on: StableSR DDPM PAL
Closest papers by meaning
- DREAM: Diffusion Rectification and Estimation-Adaptive Models (CVPR 2024 · similarity 0.88)
- BUFF: Bayesian Uncertainty Guided Diffusion Probabilistic Model for Single Image Super-Resolution (AAAI 2025 · similarity 0.88)
- When Latents Forget Pixels: Restoring Fidelity in Diffusion Transformer Super-Resolution (NeurIPS 2026 · similarity 0.88)
- SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.88)
- SkipDiff: Adaptive Skip Diffusion Model for High-Fidelity Perceptual Image Super-resolution (AAAI 2024 · similarity 0.88)
- One Diffusion Step to Real-World Super-Resolution via Flow Trajectory Distillation (ICML 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.