Research map / Image generation / Super-resolution
Building Bridges across Spatial and Temporal Resolutions: Reference-Based Super-Resolution via Change Priors and Conditional Diffusion Model
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.
Ref-Diff injects land cover change masks into a conditional EDM-based diffusion model with change-aware SFT modules to perform large-factor (8x, 16x) reference-based super-resolution of remote sensing images.
Problem
Reference-based super-resolution of remote sensing images at large scaling factors (8x-16x) suffers from unfaithful content reconstruction and ineffective reference texture transfer, especially where land cover changed between LR and Ref images.
Approach
Use off-the-shelf land cover change masks as explicit priors inside the denoising network, decoupling semantics-guided denoising (changed areas, via SFT on LR+mask) from reference texture-guided denoising (unchanged areas, via SFT on Ref+mask) within the decoder.
Key results
- On SECOND 8x, Ref-Diff reaches LPIPS 0.2642 and FID 32.60, versus HSR-Diff's 0.2689 / 45.47.
- On CNAM-CD 16x, Ref-Diff achieves LPIPS 0.3519 and FID 45.65, improving on HSR-Diff's 0.3771 / 68.20.
- Using predicted (ChangeFormer) change masks instead of ground truth slightly degrades results (e.g., SECOND 8x LPIPS 0.2657) but still beats the compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Ref-Diff instead of the strongest baseline | RefSR · SECOND 8× | LPIPS: 0.2689 -> 0.2642 | Helped |
| Use Ref-Diff instead of the strongest baseline | RefSR · SECOND 16× | LPIPS: 0.3437 -> 0.3433 | Helped |
| Use Ref-Diff instead of the strongest baseline | RefSR · CNAM-CD 8× | LPIPS: 0.3045 -> 0.2791 | Helped |
| Use Ref-Diff instead of the strongest baseline | RefSR · CNAM-CD 16× | LPIPS: 0.3771 -> 0.3519 | Helped |
| Use Ref-Diff instead of the strongest baseline | RefSR · SECOND 8× | FID: 34.5802 -> 32.5961 | Helped |
| Use Ref-Diff instead of the strongest baseline | RefSR · SECOND 16× | FID: 37.5573 -> 33.9690 | Helped |
| Use Ref-Diff instead of the strongest baseline | RefSR · CNAM-CD 8× | FID: 53.1511 -> 43.0152 | Helped |
| Use Ref-Diff instead of the strongest baseline | RefSR · CNAM-CD 16× | FID: 59.4250 -> 45.6511 | Helped |
| Add the Ref image condition (ablation) | LR-conditioned diffusion model · SECOND 8× | LPIPS: 0.3115 -> 0.2886 | Helped |
| Add the land cover change mask condition (ablation) | LR + Ref-conditioned diffusion model · SECOND 8× | LPIPS: 0.2886 -> 0.2785 | Helped |
| Add Ref texture-guided SFT (ablation) | LR + Ref + change-mask model · SECOND 8× | LPIPS: 0.2785 -> 0.2709 | Helped |
| Add semantics-guided SFT (ablation) | LR + Ref + change-mask model · SECOND 8× | LPIPS: 0.2785 -> 0.2723 | Helped |
Surprising findings
- Ref-Diff uses ground truth change masks in training and testing for the main results; predicted masks reduce performance, notably on CNAM-CD 16x (LPIPS 0.3889).
Limitations
- Main results rely on ground-truth land cover change masks; performance drops with predicted masks.
- Misleading change masks (false positives/negatives) can introduce false textures.
- Evaluated only on two remote sensing datasets with LPIPS/FID metrics, without PSNR/SSIM.
Experimental setup
Datasets: SECOND CNAM-CD
Benchmarks: SECOND CNAM-CD
Metrics: LPIPS FID
Baselines: TTSR WTRN C2-Matching AMSA DATSR EDM* HSR-Diff
Builds on: EDM HSR-Diff SPADE SFT
Closest papers by meaning
- BUFF: Bayesian Uncertainty Guided Diffusion Probabilistic Model for Single Image Super-Resolution (AAAI 2025 · similarity 0.88)
- Taming Diffusion Prior for Image Super-Resolution with Domain Shift SDEs (NeurIPS 2024 · similarity 0.88)
- SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.87)
- Unsupervised Diffusion-Based Degradation Modeling for Real-World Super-Resolution (AAAI 2025 · similarity 0.87)
- RAP-SR: RestorAtion Prior Enhancement in Diffusion Models for Realistic Image Super-Resolution (AAAI 2025 · similarity 0.86)
- DiffusionSat: A Generative Foundation Model for Satellite Imagery (ICLR 2024 · 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.