SeaDiff: Delve into Underwater Image Generation with Symmetrical Parameter Control
Keywords: diffusion model, controllable generation, underwater
TL;DR: A novel method for precise control of underwater image appearance.
Abstract: With the advancement of diffusion models, the controllability of image generation has significantly improved. However, due to the refraction and absorption of light in water, underwater images often exhibit notable variations in luminance and color cast. This leads to challenges for generative models pre-trained on terrestrial images, as they struggle to produce underwater images with a diverse range of these variations, severely limiting the appearance diversity of generated underwater images. To address this issue, we focus on the precise control of appearance in underwater images. We model the appearance of underwater images using three attributes: luminance, dynamic range, and color cast. We propose a new method, SeaDiff, which introduces a Symmetrical Parameter Control structure to achieve precise control over the appearance of underwater images. The proposed method comprises two modules: Appearance Writer, which encodes and injects appearance attributes into the U-Net encoder, and Appearance Reader, which ensures that the generated images align with the desired appearance by analyzing the feature maps. Experimental results demonstrate that the proposed SeaDiff method significantly improves control over underwater image appearance while maintaining image quality, validating its effectiveness in underwater image generation.
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.
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: 39
Loading