MFN: Metadata-Free Real-World Noisy Image Generation

Jaekyun Ko, Dongjin Kim, Guanghui Wang, Tae Hyun Kim

Keywords: real noise modeling, prompt learning, consistency models, low-level vision

Abstract:

Real-world noise poses a significant challenge in signal processing, especially for denoising tasks.
Although end-to-end denoising approaches have achieved exceptional performance, they are constrained to scenarios with abundant noisy-clean image pairs, which can be technically challenging and resource-intensive to collect.
To address this issue, several generative methods have been developed to synthesize realistic noisy images from limited real-world datasets.
While prior studies require camera metadata during training or testing to handle various real-world noise, the absence of metadata or variations in the information across different capturing devices is common in real-world scenarios, such as medical or microscope imaging, which limits their applicability.
Thus, we aim to eliminate the need for explicit camera-related labels in both stages, enhancing applicability in real-world scenarios.
To achieve this, we propose a novel framework called the Metadata-Free Noise Model (MFN), which extracts prompt features that encode input noise characteristics and generates diverse noisy images that adhere to the distribution of the input noise.
Extensive experimental results demonstrate the superior performance of our model in real-world noise generation and denoising across various benchmark datasets.

Primary Area: applications to computer vision, audio, language, and other modalities

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: 10129