Zero-shot Text-based Personalized Low-Light Image Enhancement with Reflectance Guidance
Keywords: Low-light image enhancement, Retinex decomposition, Zero-shot learning, Generative diffusion prior
TL;DR: This paper introduces a zero-shot personalized low-light image enhancement model that integrates Retinex domain knowledge into a pre-trained diffusion model, enabling style customization based on user preferences specified through text instructions.
Abstract: Recent advances in zero-shot low-light image enhancement have largely benefited from the deep image priors encoded in network architectures. However, these models require optimization from scratch for each image and cannot provide personalized results based on user preferences.
In this paper, we propose a training-free zero-shot personalized low-light image enhancement model that integrates Retinex domain knowledge into a pre-trained diffusion model, enabling style personalization based on user preferences specified through text instructions. Our contributions are as follows:
First, we incorporate the total variation optimization into a single Gaussian convolutional layer, enabling zero-shot Retinex decomposition. Second, we introduce the Contrastive Language-Image Pretraining (CLIP) model into the reflectance-conditioned sampling process of Denoising Diffusion Implicit Models (DDIM), guiding the enhancement according to user-provided text instructions. Third, to ensure consistency in content and structure, we employ patch-wise DDIM inversion to find the initial noise vector and use the reflectance as a condition during the reverse sampling process.
Our proposed model, RetinexGDP, supports any image size and produces noise-suppressed results without imposing extra noise constraints. Extensive experiments across nine low-light image datasets show that RetinexGDP achieves performance comparable to state-of-the-art models.
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
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.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
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: 143
Loading