A Fast Sampling Method in Diffusion-Based Dance Generation Models

Recently, diffusion models have attracted much attention and have also been used in many fields, including dance movement generation. However, the slow generation speed of previous sampling methods makes the diffusion-based dance generation models limited in many application scenarios. In this paper, we use a more advanced algorithm to speed up the generation of a diffusion model based system to generate long dance movements. The algorithm does not require retraining the model. Instead, it only needs to map the noise schedule of the existing model to a new time step sequence and then construct a second-order solver for the data prediction diffusion ODEs based on the estimated higher-order differentials using multi-step methods. In order to apply this algorithm to generate long sequences, we employ a technique similar to inpainting, continuously updating the correlations between short sequences during the iteration process, and eventually concatenating multiple short sequences to form a longer one. Experimental results show that our improved sampling method not only makes the generation speed faster, but also maintains the quality of the dance movements.