motion-diffusion-model icon indicating copy to clipboard operation
motion-diffusion-model copied to clipboard

How to generate visualization mesh real time

Open kai-lan opened this issue 1 year ago • 2 comments

Is there a way to generate body mesh based on model output at real-time? Isn't the model output compatible with SMPL paramters and we just need a formula like Vertices_new = Vertices_template + shape_param * beta + pose_param * pose because we have the pose? Why do we need python -m visualize.render_mesh, which seems to construct mesh from the output .mp4 video, and would take quite a few minutes?

Maybe somewhere I understood wrong?

kai-lan avatar May 02 '24 04:05 kai-lan

The SMPL mesh is currently constructed from the HumanML pose locations according to https://github.com/EricGuo5513/HumanML3D , hence there is a SMPLlify optimization in the loop that is not real-time. It should be possible to extract SMPL from the HumanML rotations (with some tweaks since they don't have the same rest pose), this should make the mesh visualization run in real-time.

GuyTevet avatar May 07 '24 08:05 GuyTevet

But SMPLify only converts joints to SMPL. Which tool can I use to obtain axis-angle data or directly convert to SMPL-X vertex?

mccuba avatar Aug 11 '24 01:08 mccuba