MoMask

MoMask is a novel masked modeling framework for text-driven 3D human motion generation, developed by researchers at the University of Alberta and Google Research and published at CVPR 2024. It employs a hierarchical quantization scheme to represent human motion as multi-layer discrete motion tokens with high-fidelity details. A Masked Transformer predicts randomly masked base-layer motion tokens conditioned on text input during training and iteratively fills in missing tokens during inference. A Residual Transformer progressively predicts next-layer tokens based on current layer results. MoMask achieves state-of-the-art performance on text-to-motion generation, with an FID of 0.045 on HumanML3D (vs 0.141 for T2M-GPT) and 0.228 on KIT-ML (vs 0.514). It also supports text-guided temporal inpainting (e.g., inbetweening, prefix, suffix) without requiring additional fine-tuning.