Research map / Image generation / Super-resolution
DifIISR: A Diffusion Model with Gradient Guidance for Infrared 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 DifIISR, a ResShift-based infrared image super-resolution diffusion model that injects gradients of FFT-spectrum (visual) and VGG/SAM-based (perceptual) losses into the noise during reverse sampling.
Problem
Infrared image super-resolution with diffusion models ignores infrared-specific thermal frequency characteristics and machine perception needs (detection and segmentation), so visually pleasing results often do not help downstream tasks.
Approach
Applies loss-gradient guidance (posterior-sampling style) to the diffusion noise prediction, with a visual loss matching normalized log-magnitude FFT spectra of HR and SR images and a perceptual loss combining VGG-16 feature MSE and SAM mask MSE.
Key results
- On M3FD (Set5), DifIISR reaches CLIP-IQA 0.6144 and MUSIQ 55.194, above SinSR (0.5877, 54.355) and ResShift (0.4701, 50.769).
- Ablation on detection: combining visual and perceptual guidance gives mAP 33.1 and mIoU 42.4 versus 31.2 and 40.9 without guidance.
- Among diffusion baselines on M3FD reference metrics (Set5), DifIISR gets PSNR 32.279 and SSIM 0.8637, ahead of SinSR and Bi-DiffSR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add visual and perceptual guidance (ablation) | DifIISR · M3FD infrared object detection | mAP: 31.2 -> 33.1 | Helped |
| Add visual and perceptual guidance (ablation) | DifIISR · FMB segmentation | mIoU: 40.9 -> 42.4 | Helped |
| Add visual guidance without perceptual guidance (ablation) | DifIISR · M3FD infrared SR | PSNR: 33.466 -> 34.528 | Helped |
| Add perceptual guidance without visual guidance (ablation) | DifIISR · M3FD infrared object detection | mAP: 31.2 -> 32.8 | Helped |
| Add both visual and perceptual guidance (ablation) | DifIISR · M3FD infrared SR | PSNR: 33.466 -> 34.575 | Helped |
| Use the gradient of the guidance loss in the noise instead of directly adding the loss (ablation) | DifIISR guidance · M3FD infrared SR | PSNR | Helped |
Surprising findings
- Reported no-reference metrics (CLIP-IQA, MUSIQ) for DifIISR exceed those of the HR images, while reference metrics lag traditional methods.
Limitations
- Reference-based metrics (PSNR, SSIM, LPIPS) lag traditional CNN/Transformer methods because outputs deviate from HR images.
- Gains on downstream tasks rely on a single detector (YOLOv5-s) and single segmenter (SegFormer-b1).
Experimental setup
Datasets: M3FD RoadScene TNO FMB
Benchmarks: M3FD RoadScene TNO FMB
Metrics: CLIP-IQA MUSIQ PSNR SSIM LPIPS mAP mIoU
Baselines: ESRGAN RealSR-JPEG BSRGAN SwinIR RealESRGAN HAT DAT ResShift CoRPLE Bi-DiffSR SinSR
Builds on: ResShift DDPM DDIM Loss-gradient guidance (Chung et al. posterior sampling) SAM
Closest papers by meaning
- One Diffusion Step to Real-World Super-Resolution via Flow Trajectory Distillation (ICML 2025 · similarity 0.86)
- DREAM: Diffusion Rectification and Estimation-Adaptive Models (CVPR 2024 · similarity 0.86)
- When Latents Forget Pixels: Restoring Fidelity in Diffusion Transformer Super-Resolution (NeurIPS 2026 · similarity 0.85)
- Uncertainty-guided Perturbation for Image Super-Resolution Diffusion Model (CVPR 2025 · similarity 0.85)
- SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.85)
- Enhancing Infrared Vision: Progressive Prompt Fusion Network and Benchmark (NeurIPS 2025 · similarity 0.85)
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.