FADI: A Fully Asymmetric Dual-Stream Interactive Network for Compact Single Image Reflection Separation

Dual-stream interactive networks have emerged as a powerful paradigm for Single Image Reflection Separation (SIRS), yet their reliance on symmetric architectures results in large models with high computational complexity. This paper challenges the necessity of this symmetric paradigm. We argue that such a design is suboptimal, motivated by the task's primary objective of transmission recovery and our empirical finding that the encoder, not the decoder, is the critical stage for feature separation. Based on these insights, we introduce FADI, a Fully Asymmetric Dual-stream Interactive network that tailors its architecture to these asymmetries for a superior performance-per-parameter trade-off. FADI's design is threefold: 1) asymmetric backbones featuring a powerful Transformer for the transmission stream and a lightweight custom block for the reflection stream; 2) a stage-aware interaction strategy with a dense module for the encoder and a sparse refinement module for the decoder; and 3) a Semantic Bootstrap process for explicit initial guidance. Extensive experiments show that FADI surpasses previous methods on multiple real-world benchmarks while being significantly more compact, establishing a new state-of-the-art for compact, high-performance SIRS.