Bilevel Layer-Positioning LoRA for Real Image Dehazing

2026-03-11

AI Summary

This work addresses the limited generalization of existing unsupervised real-world image dehazing methods under complex haze conditions and the high cost of full-model fine-tuning. The authors reformulate dehazing as a semantic alignment problem in latent space and propose a reference-free, cross-modal unsupervised guidance mechanism. Specifically, they leverage CLIP’s text–image alignment capability to construct a haze-to-clear textual guidance loss and introduce BiLaLoRA (Bilevel Localization LoRA), a novel dual-level localized adaptation module that automatically identifies optimal injection layers to jointly optimize both adapter parameters and their placement. The proposed method achieves state-of-the-art performance across multiple real-world dehazing benchmarks, with code publicly released.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Unsupervised & Self-Supervised Learning
• Natural Language Processing: Sentence-level Semantics, Textual Inference, etc.

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Learning-based real image dehazing methods have achieved notable progress, yet they still face adaptation challenges in diverse real haze scenes. These challenges mainly stem from the lack of effective unsupervised mechanisms for unlabeled data and the heavy cost of full model fine-tuning. To address these challenges, we propose the haze-to-clear text-directed loss that leverages CLIP's cross-modal capabilities to reformulate real image dehazing as a semantic alignment problem in latent space, thereby providing explicit unsupervised cross-modal guidance in the absence of reference images. Furthermore, we introduce the Bilevel Layer-positioning LoRA (BiLaLoRA) strategy, which learns both the LoRA parameters and automatically search the injection layers, enabling targeted adaptation of critical network layers. Extensive experiments demonstrate our superiority against state-of-the-art methods on multiple real-world dehazing benchmarks. The code is publicly available at https://github.com/YanZhang-zy/BiLaLoRA.

Problem

• real image dehazing
• unsupervised learning
• model adaptation
• haze removal
• computational cost

Innovation

• BiLaLoRA
• unsupervised dehazing
• CLIP-guided loss
• layer-positioning
• real image dehazing