search

Mael-zys / T2M-GPT

(CVPR 2023) Pytorch implementation of “T2M-GPT: Generating Human Motion from Textual Descriptions with Discrete Representations”
https://mael-zys.github.io/T2M-GPT/
Apache License 2.0
598 stars 52 forks source link

One question about VQ-VAE #59

Closed smandava98 closed 8 months ago

smandava98 commented 1 year ago

Hi,

Is it possible to just use the encoder to encode the motion and just regress the SMPLs directly without going through the decoder?

Say I just want to generate a couple seconds of motion. I think I am losing some information with the decoder I want to be able regress directly. Please let me know if this is possible

spzhuang commented 1 year ago

It can‘t without decoder. Because SMPL need 3D location (or 2D keypoints from a keypoints detection net) to regress it parameters, and decoder gives motion location.