The method fine-tunes a VACE branch with hand-overlay rendering while keeping the base DiT frozen. A lightweight dual-branch decoder reads a shared intermediate VACE feature to predict MANO pose, 2D joints, and translation, allowing a single inference pass to produce hand-aware reconstruction outputs without repeatedly decoding separate representations.
ViDiHand is useful for hand-object interaction analysis, robotics perception, augmented reality, motion capture, and embodied-AI datasets. Its project page includes synchronized multi-view playback, mesh and joint visualizations, comparison cases, and results on ARCTIC, HOT3D, HOI4D, and in-the-wild sequences.
