Abstract

Hand shadow art creatively uses hand shadows to produce expressive shapes on walls. We study the inverse problem: given a target shape, we aim to find hand poses that create a shadow resembling the input. This is challenging due to the vast design space of 3D hand poses and anatomical constraints. Our Hand-Shadow Poser pipeline decouples anatomical constraints from shadow shape requirements through three stages: a generative hand assignment module exploring diverse hand shape hypotheses, a hand-shadow alignment module inferring coarse poses, and a shadow-feature-aware refinement module optimizing for physical plausibility. Our approach is trainable on generic public hand data without specialized datasets. We validate our method on a benchmark of 210 diverse shadow shapes and demonstrate that it effectively generates hand poses for a wide variety of shapes in over 85% of benchmark cases.