OpenMotionLab/MotionGPT

Leverages discrete vector quantization to convert 3D motion into motion tokens—treating human movement as a "foreign language" vocabulary—enabling unified language modeling across motion and text modalities. Supports four key tasks: text-to-motion generation, motion captioning, motion prediction, and motion in-betweening through a three-stage pipeline (tokenizer pre-training, motion-language pre-training, and prompt-based instruction tuning). Integrates with HuggingFace for model distribution and builds on SMPL/HumanML3D datasets, with a PyTorch 2.0 implementation and web UI for interactive inference.

1,890 stars. No commits in the last 6 months.