The Impact of VAE Design on Latent Pose Representations for Diffusion-based Sign Language Production

This study examines the impact of variational autoencoder (VAE) design on latent pose representations for diffusion-based sign language production. The authors analyze how architectural choices and training objectives influence the latent space structure and subsequently affect the performance of a latent diffusion model, revealing that variations in generative performance, assessed via back-translation BLEU scores, are often more closely linked to latent space properties than to VAE reconstruction accuracy. This insight is crucial for practitioners as it suggests that optimizing latent space characteristics may enhance the efficacy of text-to-sign generation models.