Research map / Image generation / Image restoration
Learning Hazing to Dehazing: Towards Realistic Haze Generation for Real-World Image Dehazing
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 HazeGen, a diffusion-prior hazy image generator that produces realistic training data, and DiffDehaze, a diffusion dehazing model with an accelerated fidelity-preserving sampler (AccSamp) using AlignOp.
Problem
Real-world image dehazing depends on pre-trained models and synthetic data with poor realism, and diffusion-based dehazing is slow and loses fidelity under dense haze.
Approach
Generate realistic hazy images from a pre-trained text-to-image diffusion prior via IRControlNet with hybrid conditional/unconditional training and blended sampling, then train a diffusion dehazer whose sampler uses tiled mean/variance alignment of an early prediction to skip steps and a density-aware fidelity guidance.
Key results
- On RTTS, DiffDehaze reaches Q-Align 2.8340 and LIQE 3.0693, a 23.8% Q-Align gain over the best prior method RIDCP (2.4844).
- DiffDehaze attains best ManIQA (0.3661), CLIPIQA (0.4263), MUSIQ (65.086) and BRISQUE (16.4924) on RTTS, but FADE is worse (1.138 vs 0.712 for PTTD).
- Ablation: removing HazeGen (training DiffDehaze on synthetic data) drops Q-Align from 2.8340 to 2.1660 and LIQE from 3.0693 to 2.0014.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use hybrid conditional and unconditional training (ablation) | HazeGen · RTTS | LIQE: 2.7754 -> 3.0693 | Helped |
| Use blended sampling (ablation) | HazeGen · RTTS | LIQE: 2.6410 -> 3.0693 | Helped |
| Use hybrid training and blended sampling (ablation) | HazeGen · RTTS | LIQE: 1.5983 -> 3.0693 | Helped |
| Train DiffDehaze on HazeGen-generated data rather than synthetic data (ablation) | DiffDehaze · RTTS | LIQE: 2.0014 -> 3.0693 | Helped |
| Use AlignOp (ablation) | DiffDehaze AccSamp · RTTS | Q-Align: 2.5645 -> 2.8340 | Helped |
| Use haze density weighting in fidelity guidance (ablation) | DiffDehaze AccSamp · RTTS | Q-Align: 2.8067 -> 2.8340 | Helped |
| Use AlignOp and haze density weighting (ablation) | DiffDehaze AccSamp · RTTS | Q-Align: 2.5079 -> 2.8340 | Helped |
| Use DiffDehaze instead of PTTD | Real-world image dehazing · RTTS | FADE: 0.712 -> 1.138 | Hurt |
| Use DiffDehaze instead of RIDCP | Real-world image dehazing · RTTS | Q-Align: 2.4844 -> 2.8340 | Helped |
Surprising findings
- FADE scores flawed outputs from RIDCP and PTTD better than the proposed method, which the authors attribute to FADE's insensitivity to color-distorted haze residuals.
Limitations
- Existing no-reference dehazing metrics such as FADE are unreliable for evaluating dehazing quality.
- AccSamp's fidelity could be further improved.
Experimental setup
Datasets: RESIDE URHI RIDCP synthetic pairs RTTS Fattal's dataset
Benchmarks: RTTS Fattal's dataset
Metrics: FADE Q-Align LIQE CLIPIQA ManIQA MUSIQ BRISQUE
Baselines: DAD PSD D4 RIDCP PTTD MSBDN NAFNet spaced DDPM
Builds on: DiffBIR Stable Diffusion AdaIN classifier-free guidance RIDCP
Closest papers by meaning
- Exploiting Diffusion Prior for Real-World Image Dehazing with Unpaired Training (AAAI 2025 · similarity 0.92)
- PromptHaze: Prompting Real-world Dehazing via Depth Anything Model (AAAI 2025 · similarity 0.91)
- PRISM: Rethinking Atmospheric Scattering Reconstruction as a Unified Understanding and Restoration Model for Real-world Dehazing (NeurIPS 2026 · similarity 0.89)
- Real-world Image Dehazing with Coherence-based Pseudo Labeling and Cooperative Unfolding Network (NeurIPS 2024 · similarity 0.89)
- Tokenize Image Patches: Global Context Fusion for Effective Haze Removal in Large Images (CVPR 2025 · similarity 0.89)
- Adaptive Dynamic Dehazing via Instruction-Driven and Task-Feedback Closed-Loop Optimization for Diverse Downstream Task Adaptation (AAAI 2026 · similarity 0.88)
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.