Research map / Vision-language and multimodal / Multimodal LLMs

Adaptive Dynamic Dehazing via Instruction-Driven and Task-Feedback Closed-Loop Optimization for Diverse Downstream Task Adaptation

AAAI 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks · 1 citations (OpenAlex)

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 ADeT-Net, an adaptive dehazing network whose decoder is modulated at inference by downstream-task feedback (TFGA) and BERT-encoded text instructions (IGM), trained on ADE20K/COCO/KITTI haze synthetics and evaluated on segmentation, detection, and depth.

Problem

Standard dehazing outputs are static and not tailored to diverse downstream vision tasks; existing task-driven dehazing methods need retraining per task and lack flexibility after deployment.

Approach

A closed loop in which downstream-task outputs on the initial dehazed image and BERT-encoded user instructions jointly generate modulation weights for the decoder features of a frozen Transformer U-Net dehazing network, so the output adapts without retraining the dehazer or task models.

Key results

• On ADE20K, the proposed method reports PSNR 27.47 and SSIM 0.9701 versus 27.21 and 0.9207 for RIDCP (runner-up on PSNR).
• On Setting 1 downstream tasks, the method reports mIoU 50.34 vs 46.92 for DCMPNet and mAP50-95 35.7 vs 34.2 for DCMPNet.
• Under Setting 2 (no adaptation), the method reports mIoU 50.34, comparable to its Setting 1 numbers, while the baselines range from 39.76 to 45.35 mIoU.

Surprising findings

• Setting 2/3 tables for the proposed method repeat the exact Setting 1 numbers, which is unexpected if the closed-loop modulation changes the dehazed outputs at inference.

Limitations

• Evaluated on only a fixed set of three downstream tasks (segmentation, detection, depth).
• Adaptation to dynamically changing task types and requirements is not addressed.
• Baseline comparison uses task networks retrained per method, while the proposed method's downstream numbers are not clearly distinguished across settings.

Experimental setup

Datasets: ADE20K COCO KITTI

Benchmarks: ADE20K COCO KITTI

Metrics: PSNR SSIM LPIPS mIoU mAP mAP50-95 AbsRel SqRel RMSE RMSElog

Baselines: Dehamer C2P RIDCP Dehazeformer KA-Net DCMPNet DiffDehaze IPC

Builds on: U-Net SegFormer YOLOv5 RADepth BERT

Closest papers by meaning

• Learning Hazing to Dehazing: Towards Realistic Haze Generation for Real-World Image Dehazing (CVPR 2025 · similarity 0.88)
• Depth Information Assisted Collaborative Mutual Promotion Network for Single Image Dehazing (CVPR 2024 · similarity 0.86)
• PromptHaze: Prompting Real-world Dehazing via Depth Anything Model (AAAI 2025 · similarity 0.86)
• Tokenize Image Patches: Global Context Fusion for Effective Haze Removal in Large Images (CVPR 2025 · similarity 0.85)
• Iterative Predictor-Critic Code Decoding for Real-World Image Dehazing (CVPR 2025 · similarity 0.85)
• PRISM: Rethinking Atmospheric Scattering Reconstruction as a Unified Understanding and Restoration Model for Real-world Dehazing (NeurIPS 2026 · similarity 0.85)