Research map / Image generation / Image restoration
Exploiting Diffusion Prior for Real-World Image Dehazing with Unpaired Training
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 Diff-Dehazer, an unpaired real-world image dehazing framework that puts SD Turbo diffusion backbones inside a CycleGAN cycle, with text-aware and physics-aware (DCP/BCCR) guidance.
Problem
Real-world image dehazing with unpaired training: existing unpaired methods (CycleGAN-style) have limited feature representation, and diffusion-based dehazing relies on synthetic paired data, so generalization to real scenes is limited.
Approach
Uses a distilled one-step diffusion model (SD Turbo) as bijective hazing/dehazing mappers inside a CycleGAN cycle, adds a physics-aware loss that reconstructs hazy images via the atmospheric scattering model from DCP and BCCR outputs fused by a Perceptual Fusion Model, and adds text-aware guidance using BLIP-2 captions with haze terms removed plus a learned textual-inversion haze prompt.
Key results
- On RTTS, the method reaches FID 52.344 vs 65.076 for the best prior baseline (RefineDNet), and on Haze2020 FID 71.984 vs 80.965 (Diff-Plugin).
- On OHAZE, PSNR 19.539 and LPIPS 0.195, best among compared methods (next best PSNR 18.693 for RefineDNet).
- Ablation: removing TAG and PAG raises FID from 71.984 to 74.575 on Haze2020.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add PAG to the backbone without TAG (ablation) | Diff-Dehazer · Haze2020 | FID: 74.575 -> 73.096 | Helped |
| Add TAG to the backbone without PAG (ablation) | Diff-Dehazer · Haze2020 | FID: 74.575 -> 74.100 | Helped |
| Add TAG to the backbone with PAG (ablation) | Diff-Dehazer · Haze2020 | FID: 73.096 -> 71.984 | Helped |
| Add PAG to the backbone with TAG (ablation) | Diff-Dehazer · Haze2020 | FID: 74.100 -> 71.984 | Helped |
| Add PAG to the model without PAG (ablation) | Diff-Dehazer · Haze2020 | FID: 74.100 -> 76.768 | Hurt |
| Add BCCR to the model without PAG (ablation) | Diff-Dehazer · Haze2020 | FID: 74.100 -> 74.999 | Hurt |
| Use both DCP and BCCR rather than either alone (ablation) | Diff-Dehazer · Haze2020 | FID: 76.768 -> 71.984; 74.999 -> 71.984 | Helped |
| Add a positive text prompt (ablation) | Diff-Dehazer · real-world image dehazing | dehazing effect | Helped |
| Add a negative text prompt alongside the positive prompt (ablation) | Diff-Dehazer · real-world image dehazing | dehazing effect | Helped |
| Use Diff-Dehazer rather than KANet | real-world dehazing · RTTS | FID: 64.963 -> 52.344 | Helped |
| Use Diff-Dehazer rather than C2PNet | real-world dehazing · RTTS | CLIPIQA: 0.390 -> 0.328 | Hurt |
| Use Diff-Dehazer rather than Diff-Plugin | real-world dehazing · Haze2020 | FID: 80.965 -> 71.984 | Helped |
Surprising findings
- The method has lower CLIPIQA than C2PNet on RTTS (0.328 vs 0.390), which the authors attribute to physical priors limiting diffusion stochasticity.
Limitations
- Fine-tuned diffusion may produce diversified outputs and misalignment in cases with severe haze.
- Relies on physical prior methods (DCP, BCCR) as sources of physical supervision.
Experimental setup
Datasets: RESIDE (URHI, RTTS subsets) ADE20K OTS
Benchmarks: RTTS Haze2020 OHAZE
Metrics: FID NIQE MUSIQ CLIPIQA PSNR SSIM LPIPS VSI
Baselines: DCP BCCR RefineDNet PSD Dehamer D4 DehazeFormer C2PNet KANet InstructIR Diff-Plugin
Builds on: CycleGAN SD Turbo Atmospheric Scattering Model classifier-free guidance textual inversion
Closest papers by meaning
- Learning Hazing to Dehazing: Towards Realistic Haze Generation for Real-World Image Dehazing (CVPR 2025 · similarity 0.92)
- PromptHaze: Prompting Real-world Dehazing via Depth Anything Model (AAAI 2025 · similarity 0.89)
- DeblurDiff: Real-Word Image Deblurring with Generative Diffusion Models (NeurIPS 2025 · similarity 0.89)
- CoA: Towards Real Image Dehazing via Compression-and-Adaptation (CVPR 2025 · similarity 0.88)
- Diffusion Priors for Variational Likelihood Estimation and Image Denoising (NeurIPS 2024 · similarity 0.88)
- Prior-guided Hierarchical Harmonization Network for Efficient Image Dehazing (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.