CObL: Toward Zero-Shot Ordinal Layering Without User Prompting

Figure 1. Our model, CObL, infers a stack of occlusion-ordered object layers that composite back to the image. (Left) We train CObL using 2250 synthetic 3D tabletop scenes created using 3D shapes with generative textures and lighting. (Right) Once trained, CObL generalizes to captured photographs of tabletops with variable numbers of novel objects. It concurrently generates a stack of amodally-completed object layers while using inference-time guidance to ensure the layers composite to the input.