RefMover: Diffusion-Based Single Image Reflection Removal With Language and Region Guidance.
TL;DR
RefMover is introduced, a multi-instruction diffusion framework for SIRR that bridges the semantic-spatial gap by combining language and region guidance and features a pipeline with a streamlined conditional architecture that mitigates condition conflicts and preserves details.
Abstract
Single image reflection removal (SIRR) is a highly ill-posed task. While recent research explores RAW data for physical accuracy, the dominance of processed (8-bit) images in real-world scenarios necessitates robust solutions capable of handling non-linear degradations. Prevailing methods, however, can struggle on such data due to the lack of high-level semantic guidance. Although language-guided diffusion models show promise in bridging this gap, they face challenges in semantic-spatial alignment and recovery faithfulness due to contaminated input conditions. To address these challenges, this paper introduces RefMover, a multi-instruction diffusion framework for SIRR. It bridges the semantic-spatial gap by combining language and region guidance. To improve faithful and stable recovery, RefMover features a pipeline with a streamlined conditional architecture that mitigates condition conflicts and preserves details. We also introduce a new large-scale dataset (ComRR) for robust training and evaluation. The experimental results demonstrate the effectiveness of the proposed method and its potential for practical interactive and cloud-assisted applications.