EDGE: Editable Dance Generation From Music

### Model/Pipeline/Scheduler description

The authors of this paper propose a diffusion-based method for generating human dance sequences (represented by joint angle trajectories) conditioned on music audio embeddings. The weights are stored in Google Docs and are open to the public, but translating the weights into a Diffusers-compatible format will require some surgery. Incorporating this model may also require implementing FiLM timestep conditioning and other possibly currently unavailable NN modules in Diffusers.

### Open source status

- [X] The model implementation is available
- [X] The model weights are available (Only relevant if addition is not a scheduler).

### Provide useful links for the implementation

Paper: https://arxiv.org/pdf/2211.10658.pdf
Website: https://edge-dance.github.io/
Github: https://github.com/Stanford-TML/EDGE

@jtseng20 @Stanford-TML

Start with the linked EDGE GitHub repository and paper to understand the model implementation, then inspect the publicly available weights referenced in the issue. Determine the work needed to translate the weights into a Diffusers-compatible format and whether FiLM timestep conditioning or other unavailable modules are required. Done means the EDGE model is incorporated with its required components and can generate dance sequences conditioned on music.

Vom Indexierungsmodell aus dem Issue-Text verfasst.