Text-Conditioned Motion Generation

Advisor: Wangmeng Zuo, Professor at Harbin Institute of Technology

• The Hydra library is used for the parsing, and the dataset splits utilize the same splits as in Language2Pose (train /Val / test).
• Redefining the parameters that fit the SMPL dataset to the model resulted in the creation of a new dataset.
• By combining several motion-sequence embeddings using transformers, it is possible to encode the VAE’s distribution parameters as in.
• Sampling from a learned distribution function of motion lengths conditional on the input text, we subsequently use a temporal variant autoencoder to synthesize a set of human actions of different lengths.
• Use the diffusion model to enhance the quality of the generated animation when generating human animation.