ReMoMask: Retrieval-Augmented Masked Motion Generation
The official code for an ECCV 2026 paper that generates realistic human motion sequences from text descriptions by combining retrieval of similar existing motions with a masked transformer generation model.
ReMoMask addresses two separate weaknesses in prior text-to-motion approaches: generative models such as diffusion, which tend to produce limited variety, accumulate error over a sequence, and can generate physically implausible motion; and retrieval-augmented approaches, which the paper describes as suffering from diffusion inertia, partial mode collapse and asynchronous artifacts. It introduces three components together: a bidirectional momentum text-motion model that decouples negative sample scale from batch size to improve retrieval precision, a semantic spatiotemporal attention mechanism that enforces biomechanical constraints during part-level fusion, and a retrieval-augmented classifier-free guidance step that adds minor unconditional generation to improve generalization.
Built on top of the MoMask model's residual vector-quantized VAE, it generates temporally coherent motion in a small number of steps, and the paper reports state-of-the-art results on standard benchmarks including HumanML3D. A successor, ReMoMask-2, is also referenced as adding latent-aligned retrieval for stronger, single-stage generation.
It is written in Python; the paper's license is not asserted in the repository metadata.