We present a real-time method for multiview synthesis from a single input image. Our approach predicts a depth-peeling–like layered scene representation directly from the image, enabling view-dependent reconstruction without explicit 3D geometry. Pr...
We present a real-time method for multiview synthesis from a single input image. Our approach predicts a depth-peeling–like layered scene representation directly from the image, enabling view-dependent reconstruction without explicit 3D geometry. Previous layer-based approaches attempt to recover full scene geometry or hallucinate entire backgrounds, which are computationally expensive for real-time applications. We define the concept of virtually visible hidden volumes (VVHVs) that geometrically delimit the hidden regions in the source view that become visible in the bounded target views, serving as an effective visibility domain for reconstruction.The layered representation is inferred using a kernel-predicting neural network, which reconstructs occluded regions as a weighted aggregation of visible areas in the input view. Furthermoe, we employ a multi-view image-warping scheme optimized for both the layered scene representation and the bounded set of novel views, enabling fast and consistent multi-view synthesis. Our real-time solution achieves image quality comparable to that of offline single-image novel view synthesis methods, demonstrating both high efficiency and visual fidelity.