OmniClear: Soft Effects Removal from Images within a Versatile Model

Jingdong Zhang, Lingzhi Zhang, Qing Liu, Mang Tik Chiu, Connelly Barnes, Yizhou Wang, Haoran You, Xiaoyang Liu, Yuqian Zhou, Zhe Lin, Eli Shechtman, Sohrab Amirghodsi, Xin Li, Wenping Wang, Xiaohang Zhan

Keywords: Image Editing, Low-level Vision

TL;DR: Proposing a versatile model removing multiple soft effects degrading image quality in photography.

Abstract: Digital images are often degraded by soft effects such as lens flare, haze, shadows, and reflections, which reduce aesthetics even though the underlying pixels remain partially visible. The prevailing works address these degradations in isolation, developing highly specialized, specialist models that lack scalability and fail to exploit the shared underlying essences of these restoration problems. While specialist models are limited, recent large-scale pretrained generalist models offer powerful, text-driven image editing capabilities. while recent general-purpose systems (e.g., GPT-4o, Flux Kontext, Nano Banana) require detailed prompts and often fail to achieve robust removal on these fine-grained tasks or preserve identity of the scene. Leveraging the common essence of soft effects, i.e., semi-transparent occlusions, we introduce a medium-scale foundational versatile model, capable of addressing diverse degradations caused by soft effects within a single framework. Our approach centers on fine-tuning a potent inpainting model on a large-scale, curated dataset of paired images, enabling it to learn robust restoration priors. Our method provides simple and intuitive user control, either global removal or mask-based removal with strength control, making interaction easier while ensuring higher reliability. Extensive experiments demonstrate that our unified model outperforms both prior specialist methods and popular general-purpose models, achieving robust and stable performance on in-the-wild scenarios.

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

Submission Number: 6246