Latent-Space Generative Adversarial Networks for Text-Conditioned Human Motion Synthesis
A method for synthesizing 3D human motion from a text prompt by training a Generative Adversarial Network entirely in the latent space of a fixed motion Variational Autoencoder, with a pretrained CLIP text encoder supplying the conditioning. By generating latent codes with a simple GAN rather than a latent diffusion process, it reaches motion quality comparable to state-of-the-art diffusion models (FID 0.482 on HumanML3D) while cutting inference compute by over 90 percent. A student learns how moving a generative model into a learned latent space trades an expensive iterative sampler for a fast single-shot generator.
LS-GAN: Human Motion Synthesis with Latent-space GANs Avinash Amballa Gayathri Akkinapalli
This paper (LS-GAN) synthesizes 3D human motion from a text prompt using a Generative Adversarial Network that operates in the latent space of a fixed motion Variational Autoencoder. The VAE (a trans…