T2LM: Long-Term 3D Human Motion Generation from Multiple Sentences
Researchers have developed T2LM, a novel framework for generating long sequences of 3D human motion from multiple sentences. Unlike previous methods that required sequential motion data and often produced unrealistic gaps, T2LM can be trained without such data. It utilizes a VQ-VAE to compress motion into latent vectors and a Transformer-based text encoder to predict these vectors from text, enabling smoother transitions and improved motion generation. AI
IMPACT Enables more realistic and extended 3D character animations for gaming, film, and virtual environments.