question about idle motion

[cyk990422]

I would like to understand why, when the character is supposed to remain standing, the generated digital-human motion exhibits floating and large spatial drifting. The model generates idle motions quite well visually, but these motions seem to accumulate joint or pose errors over time. What could be the underlying reason for this?

[caiyy17]

That's a good question. In this work, we train the model purely with the standard diffusion loss, without any additional tricks. As a result—like many streaming models—the generated motion suffers from distribution drift.

At the start, the error is small (say, $\Delta x$). The model then generates the next motion conditioned on this slightly erroneous state (which is the history part), which compounds the error (to $\sim 2\Delta x$), and so on. The error keeps growing until it eventually saturates at some level. This is what produces the floating and spatial drift you observe.

This problem can be mitigated by exposing the model to its own prediction errors during training—e.g., via similar technics like self-forcing.