PAMD: Plausibility-Aware Motion Diffusion Model for Long Dance Generation

Ying Zhu*,

> Department of Computer Science and Engineering, Southeast University, Nanjing, China

> Key Laboratory of New Generation Artificial Intelligence Technology and Its Interdisciplinary Applications

We introduce PAMD, a framework for generating dances that are both musically aligned and physically realistic. Here, we present 10 pairs of comparison videos with the ground truth data from the AIST++ test set, which includes 2 pairs with a duration of 5 seconds, 5 pairs with a duration of 7.5 seconds, and 3 pairs with a duration of 10 seconds. Blue motions are generated by EDGE, grey motions are generated by PAMD (our method) and pink motions are ground truth.

5-second dances

1.House:

EDGE

PAMD

GT

2.Ballet Jazz:

EDGE

PAMD

GT

7.5-second dances

1.Krump:

EDGE

PAMD

GT

2.Street Jazz:

EDGE

PAMD

GT

3.LA style Hip-hop:

EDGE

PAMD

GT

4.Lock:

EDGE

PAMD

GT

5.Middle Hip-hop:

EDGE

PAMD

GT

10-second dances

1.Break:

EDGE

PAMD

GT

2.Pop:

EDGE

PAMD

GT

3.Waack:

EDGE

PAMD

GT