FreeDehaze: Towards Training-free Real-world Image Dehazing via Diffusion Degradation Prior

Restoring high-quality images from degraded hazy images is a challenging task, particularly in real-world scenarios where the complexity and diversity of haze limit the generalization of existing methods. Recent investigations seek to address this limitation by exploring advanced methods for synthesizing haze and incorporating real-world hazy images. Due to the inherent diversity and complexity of real-world haze, these methods struggle to accurately model haze representations. In this paper, we observe that hazy images generated by advanced text-to-image diffusion models exhibit a remarkable resemblance to real-world haze, suggesting that they effectively internalize haze representations. Inspired by this observation, we propose FreeDehaze, a novel training-free method for real-world image dehazing. FreeDehaze is a posterior-based framework capable of addressing non-linear dehazing challenges without relying on additional degradation estimation networks. It follows the human cognition for image restoration, beginning with perception and subsequently enhancing the image. The core methodology of FreeDehaze involves generating pseudo-clean images by transforming abstract textual descriptions, interpreted through vision-language models, into specific reference images. Subsequently, it employs optimal transport to align the dehazing prediction with the pseudo-clean image within the haze sub-space, enabling high-fidelity dehazing. Experimental results demonstrate that FreeDehaze outperforms existing methods in subjective visual quality on challenging datasets (e.g., RTTS, URHI, and Fattal) and achieves competitive objective metrics, demonstrating strong generalization without the need for additional training.