Efficient Masked Decoding for Sign Motion Generation with Hand-Aware Token Refinement

Text-to-sign language motion generation has been dominated by autoregressive (AR) decoders, usually based on transformers. Such architectures incur high inference and training times. We make three contributions. First, we propose a masked iterative non-autoregressive generator that produces full-body sign sequences from free text through parallel decoding. Second, the model predicts sequence length directly from the textual input, removing the need for an external duration model. Third, we introduce a training-free, parameter-free hand-aware refinement stage that asymmetrically allocates additional computation, selectively regenerating only the least-confident hand tokens, so that articulated hand motion is improved without retraining the model. As a result, on CSL-Daily and Phoenix-2014T our method improves pose accuracy over the AR baseline by up to 13.6% on body pose and 14.8% on hand pose. It runs 1.73x to 2.02x faster end-to-end. And it consumes roughly a quarter of the GPU energy per generated sequence.