From Generation to Restoration in Single-Image Reflection Removal

Mingjia Li, Jin Hu, Qiming Hu, Jiarui Wang, Hainuo Wang, Xiaojie Guo

Keywords: Reflection Removal

Abstract: Single-image reflection removal (SIRR) is a highly ill-posed problem, where existing discriminative methods struggle to recover regions heavily corrupted by reflections and often fail to generalize in the wild. This work presents a new framework that reframes SIRR as a guided generation task by adapting a pre-trained Diffusion Transformer (DiT) into a precise restoration model. The key principle is to regulate the generative flexibility of DiTs within a structured latent space. To this end, we design two core components, including i) a reflection-equivariant VAE that encodes reflection artifacts into a compact latent prior; and ii) a set of learnable prompts that provides direct, task-specific guidance while bypassing the ambiguity of text-based conditioning; These designs transform a general-purpose image editing DiT into a precise and robust tool for reflection removal, capable of reconstructing transmission layers with high fidelity and fine detail. Extensive experiments reveal that our model achieves new state-of-the-art performance on standard benchmarks and, critically, generalizes strongly to challenging real-world images. Code will be made publicly available.

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

Submission Number: 2039