Principled Reflection Separation via Nonlinear Superposition and Feature Interaction
Two new research papers propose advanced methods for separating reflections from single images, a challenging task in computer vision. One paper introduces a diffusion model that jointly generates transmission and reflection layers, employing cross-layer attention and a disjoint sampling strategy to improve disentanglement. The second paper revisits the problem by proposing a learnable nonlinear superposition model and a dual-stream framework that captures bidirectional dependencies between layers, moving beyond simplified linear composition assumptions. AI
IMPACT Advances in reflection separation could improve image editing, autonomous driving perception, and augmented reality applications.