Reproducing Prompt Result Shown in Demo

[GaetanoDibenedetto]

I’ve installed all the required packages and tested the model. However, I noticed that the results for the same prompt differ significantly from those shown in the demo video, and they also appear less accurate. Am I doing something wrong? Could I be missing a step, such as loading a pretrained model?

For example, I used the prompt:
"Generate a man aggressively kicking an object to the left using his right foot."

Demo output from the authors:
https://github.com/user-attachments/assets/c2df2d1e-a721-47ff-a828-dba210c57bd9

Output I obtained locally:
https://github.com/user-attachments/assets/92b70874-6602-4515-8311-3615d04c1e15

[DominikVincent]

I have some similar problems, but the other direction. When using the model for motion to text the descriptions the generated descriptions don't match the motion at all. I am running on all humanML3D motions and independent of the split that I am using the motion description is very bad.

I processed the HumanML3D script following their instructions and use the generated new_joint_vecs motions as input to the model following the demo code that you provided. I verified that both the demo code and the automated eval code for whole of humanML3D provide identical results.

What configuration did you run to do motion to text?