Some issues about motion editing (motion in-betweening) evaluation without text condition

[Coondinator]

I noticed that the code doesn't include the motion editing without text and the default training doesn't use CFG. How can I achieve motion in-betweening without providing text information? I tried setting the word embeddings to zero or leaving the text input empty, but the results were extremely poor.

[exitudio]

We retrain the model without text condition for this evaluation.

[Coondinator]

We retrain the model without text condition for this evaluation.

I see. So for this task, the only conditions are the fixed keyframes(tokens) at the beginning and end?

[exitudio]

Correct.
Only text conditions change (for both training and testing) everything else remains the same.