Addressing domain shift with diffusion-based adaptation for real image dehazing
Keywords: diffusion-based adaptation, domain gap, real image dehazing
Abstract: Conventional supervised single-image dehazing methods, which are trained with substantial synthetic hazy-clean image pairs, have achieved promising performance. However, they often fail to tackle out-of-distribution hazy images, due to the domain shift between source and target scenarios (e.g., between indoor and outdoor, between synthetic and real). In this work, we observe the opportunity for improving such dehazing models' generalization ability without modifying the architectures or weights of conventional models by adopting the diffusion model to transfer the distribution of input images from target domain to source domain. Specifically, we train a denoising diffusion probabilistic model (DDPM) with source hazy images to capture prior probability distribution of the source domain. Then, during the test-time the obtained DDPM can adapt target hazy inputs to source domain in the reverse process from the perspective of conditional generation. The adapted inputs are fed into a certain state-of-the-art (SOTA) dehazing model pre-trained on source domain to predict the haze-free outputs. Note that, the whole proposed pipeline, termed \textbf{Diff}usion-based \textbf{AD}aptation (DiffAD), is model-agnostic and plug-and-play. Besides, to enhance the efficiency in real image dehazing, we further employ the predicted haze-free outputs as the pseudo labels to fine-tune the underlying model. Extensive experimental results demonstrate that our DiffAD is effective, achieving superior performance against SOTA dehazing methods in domain-shift scenarios.
Supplementary Material: zip
Primary Area: generative models
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 10523
Loading