Li Siyao

Hi, thank you for your interest! To drive character animation, you need to transfer 3D joint keypoints to rotation matrix/eular angles (this step is called "inverse kinemetics", IK), save them...

Hi, thank you for your interest! To drive character animation, you need to transfer 3D joint keypoints to rotation matrix/eular angles (this step is called "inverse kinemetics", IK), save them...

I have add links to NutCloud but you have to download the four weights separately since the NutCloud's 500M size limitation.

Yes! It is feasible to do so. I am sorry for I did not clean the code sufficiently.

Hi there! Thanks for your interests! Referring to AIST++ benchmark [1], FID values are computed using features of generated dances and ALL real dances, which includes both training and test...

Hahah! Thank you so much. I am trying to make the code cleaner to be not quite painful to read. Just few days.

I received your info. I will solve it this weekend. However, the original paper is implemented using 3D keypoint positions. So I did not intend to focus on supporting SPML...

I believe the reason is that in sep_vqvae step, the final visulaization uses the ground truth root shift (because VQVAE in this step is not trained with the global velocity...

Hi, thanks for your interesting in watching the code in detail! Actually, the mode we used is not "summation". You could find normalize below~

PyTorch1.1 is the version in the cluster we used for this project, which is not easy to update... We assumed some potential difference after 1.2 so we claimed the environment...