Research map / Image generation / Image restoration
PromptHaze: Prompting Real-world Dehazing via Depth Anything 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.
Proposes PromptHaze, a real-world image dehazing method that uses Depth Anything depth prompts updated iteratively with a controllable dehazing-strength network (NAFNet-based), trained on online-synthesized haze.
Problem
Real-world image dehazing is hard due to diverse haze degradation and lack of large-scale paired real data; existing prior-based and generative-codebook methods struggle to recover backgrounds and fine details in dense haze.
Approach
Uses depth features from the frozen Depth Anything model as prompts for a dehazing network with a controllable strength factor alpha, and iteratively refines the depth prompt by re-running dehazing on previous outputs (prompt-by-prompt strategy), with haze-parameter regression injected via AdaLN.
Key results
- On RTTS, PromptHaze achieves BRISQUE 14.624 and NIMA 5.1141, versus RIDCP at 18.782 and 4.4267.
- FADE is 0.929 for PromptHaze, second to PSD at 0.920.
- User study preference on 80 RTTS images: 51.2% for PromptHaze vs 16.2% for RIDCP (next best).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the depth prompt (ablation) | PromptHaze · RTTS | BRISQUE: 19.712 -> 14.624 | Helped |
| add the depth prompt (ablation) | PromptHaze · RTTS | NIMA: 4.1041 -> 5.1141 | Helped |
| add contrastive regularization (ablation) | PromptHaze · RTTS | BRISQUE: 16.824 -> 14.624 | Helped |
| add contrastive regularization (ablation) | PromptHaze · RTTS | NIMA: 4.5190 -> 5.1141 | Helped |
| add the Prompt-by-Prompt Dehazing Strategy (ablation) | PromptHaze · RTTS | BRISQUE: 16.791 -> 14.624 | Helped |
| add the Prompt-by-Prompt Dehazing Strategy (ablation) | PromptHaze · RTTS | NIMA: 4.8437 -> 5.1141 | Helped |
| add haze-relevant parameter projection (ablation) | PromptHaze · RTTS | BRISQUE: 15.944 -> 14.624 | Helped |
| add haze-relevant parameter projection (ablation) | PromptHaze · RTTS | NIMA: 4.8141 -> 5.1141 | Helped |
Surprising findings
- Directly feeding the Depth Anything depth map as a condition into the dehazing network did not give ideal results because haze-induced depth errors accumulate.
Limitations
- Artifacts and color distortions in foreground for challenging scenes
- Multiple inference iterations make it slow, not real-time
- Evaluation relies on non-reference metrics and user study with no paired ground truth on real data
Experimental setup
Datasets: RIDCP 500 clean images with depth maps (online haze synthesis)
Benchmarks: RTTS Fattal's dataset
Metrics: FADE BRISQUE NIMA user study preference
Baselines: MSBDN FFA-Net Dehamer DAD PSD RIDCP
Builds on: Depth Anything RIDCP NAFNet Contrastive regularization dehazing
Closest papers by meaning
- Learning Hazing to Dehazing: Towards Realistic Haze Generation for Real-World Image Dehazing (CVPR 2025 · similarity 0.91)
- SelfPromer: Self-Prompt Dehazing Transformers with Depth-Consistency (AAAI 2024 · similarity 0.90)
- Exploiting Diffusion Prior for Real-World Image Dehazing with Unpaired Training (AAAI 2025 · similarity 0.89)
- Tokenize Image Patches: Global Context Fusion for Effective Haze Removal in Large Images (CVPR 2025 · similarity 0.89)
- Dehaze-RetinexGAN: Real-World Image Dehazing via Retinex-based Generative Adversarial Network (AAAI 2025 · similarity 0.88)
- Real-world Image Dehazing with Coherence-based Pseudo Labeling and Cooperative Unfolding Network (NeurIPS 2024 · 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.