Absolute Coordinates Make Motion Generation Easy
Paper • 2505.19377 • Published • 3
Autoregressive version of ACMDM (Absolute Coordinates Motion Diffusion Model) inspired by JEPA. Trained on HumanML3D dataset, same as original model. Simulates motion token by token in latent space which can be decoded to joint positions. Time to first token is about 0.1s and each frame takes about 0.06 seconds total to generate and decode so practically real-time motion generation.
This contains both the text-to-motion checkpoint and the ControlNet version. You can find the Github repo here. Keep the existing directory structure and place it in the directory 'checkpoints/t2m' alongside the other required models. You can find the original repo in the acknowledgements which will tell you how to obtain those models as well as the dataset.
Acknowledgements:
Meng, Z., Han, Z., Peng, X., Xie, Y., & Jiang, H. (2025). Absolute Coordinates Make Motion Generation Easy. arXiv Preprint arXiv:2505. 19377.
Links to their paper here and to their code implementation here.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support