Research map / Image generation / Image restoration
Iterative Predictor-Critic Code Decoding 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 IPC-Dehaze, an iterative predictor-critic code decoding framework over a pre-trained VQGAN codebook that refines codes across iterations for real-world image dehazing.
Problem
Real-world image dehazing struggles with spatially varying haze density and depth; existing codebook-based dehazing methods use one-shot code prediction, which fails on dense haze and thin-haze detail.
Approach
Replace one-shot nearest-neighbor codebook matching with an iterative decoding loop: a Code-Predictor conditioned on previously obtained high-quality codes predicts codes, and a Code-Critic evaluates code correlations to decide which codes to retain or resample, going from easy (thin haze) to hard (dense haze) regions.
Key results
- On RTTS, IPC-Dehaze reaches MUSIQ 59.60, MANIQA 0.327, CLIPIQA 0.44, TOPIQ 0.500, the best among compared methods on no-reference IQA metrics.
- On Fattal's dataset, IPC-Dehaze obtains the top scores on MUSIQ (66.22), MANIQA (0.43), CLIPIQA (0.59), Q-Align (4.234) and TOPIQ (0.63).
- Ablation on RTTS with T=8: replacing Code-Predictor with NN matching drops MUSIQ from 59.60 to 58.19 and TOPIQ from 0.500 to 0.458.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use iterative Predictor-Critic code decoding instead of one-shot decoding | IPC-Dehaze · RTTS | MUSIQ: 55.23 -> 59.60 | Helped |
| use iterative Predictor-Critic code decoding instead of one-shot decoding | IPC-Dehaze · Fattal | MUSIQ: 65.48 -> 66.22 | Helped |
| use iterative Predictor-Critic code decoding instead of one-shot decoding | IPC-Dehaze · URHI | MUSIQ: 61.39 -> 62.5 | Helped |
| use iterative Predictor-Critic code decoding instead of one-shot decoding | IPC-Dehaze · RTTS | PI: 3.56 -> 3.22 | Helped |
| use iterative Predictor-Critic code decoding instead of one-shot decoding | IPC-Dehaze · Fattal | PI: 2.37 -> 2.41 | Hurt |
| use iterative Predictor-Critic code decoding instead of one-shot decoding | IPC-Dehaze · URHI | PI: 2.87 -> 3.08 | Hurt |
| add Code-Predictor rather than use nearest-neighbor code matching (ablation) | IPC-Dehaze · RTTS | MUSIQ: 58.19 -> 59.60 | Helped |
| add Code-Predictor rather than use nearest-neighbor code matching (ablation) | IPC-Dehaze · RTTS | PI: 3.25 -> 3.22 | Helped |
| add Code-Predictor rather than use nearest-neighbor code matching (ablation) | IPC-Dehaze · RTTS | MANIQA: 0.303 -> 0.327 | Helped |
| add Code-Predictor rather than use nearest-neighbor code matching (ablation) | IPC-Dehaze · RTTS | Q-Align: 3.25 -> 3.49 | Helped |
| add Code-Predictor rather than use nearest-neighbor code matching (ablation) | IPC-Dehaze · RTTS | TOPIQ: 0.458 -> 0.500 | Helped |
| add Code-Critic (ablation) | IPC-Dehaze · RTTS | MUSIQ: 57.74 -> 59.60 | Helped |
Surprising findings
- Nearest-neighbor code matching gives no gain from more iterations, since per-token independent matching does not change codes across iterations.
Limitations
- Evaluation relies only on no-reference IQA metrics because real hazy images lack ground truth.
- Iterative decoding requires T passes at inference, adding cost (not quantified in the text).
- Training data is synthetic, so real-world generalization depends on the degradation model.
Experimental setup
Datasets: DIV2K Flickr2K RIDCP synthetic hazy data
Benchmarks: RTTS Fattal URHI
Metrics: MUSIQ PI MANIQA CLIPIQA Q-Align TOPIQ
Baselines: MSBDN Dehamer DEA-Net DAD D4 PSD RIDCP KA-Net
Builds on: VQGAN MaskGIT RIDCP FeMaSR SFT
Closest papers by meaning
- Learning Hazing to Dehazing: Towards Realistic Haze Generation for Real-World Image Dehazing (CVPR 2025 · similarity 0.87)
- CoA: Towards Real Image Dehazing via Compression-and-Adaptation (CVPR 2025 · similarity 0.87)
- Tokenize Image Patches: Global Context Fusion for Effective Haze Removal in Large Images (CVPR 2025 · similarity 0.86)
- PRISM: Rethinking Atmospheric Scattering Reconstruction as a Unified Understanding and Restoration Model for Real-world Dehazing (NeurIPS 2026 · similarity 0.86)
- PromptHaze: Prompting Real-world Dehazing via Depth Anything Model (AAAI 2025 · similarity 0.85)
- Adaptive Dynamic Dehazing via Instruction-Driven and Task-Feedback Closed-Loop Optimization for Diverse Downstream Task Adaptation (AAAI 2026 · 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.