VGGA: Geometry-Grounded Autoencoder for Video Generation

Key points

• Autoencoders are the core component of latent video generative models, compressing high-dimensional videos into compact latent spaces for downstream generation.1
• We argue that videos should not be treated merely as signals to be reconstructed, but rather as 2D projections of dynamic 3D worlds that exhibit explicit and well-structured geometric properties.1
• Specifically, VGGA grounds the learned representations along three complementary geometric dimensions: Structural Depth, Motion Correspondence, and Projective Warping.1
• These objectives are unified within a Video Geometry Decoder, which employs a shared spatio-temporal upsampling trunk to jointly optimize geometric learning and RGB reconstruction.1

Sources (1)

• [1]VGGA: Geometry-Grounded Autoencoder for Video Generation
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 12, 05:34 AM

  Building on this insight, we propose Video Geometry Grounded Autoencoder (VGGA), a geometry-aware video autoencoder that explicitly incorporates geometric constraints into the reconstruction process.

  Autoencoders are the core component of latent video generative models, compressing high-dimensional videos into compact latent spaces for downstream generation.

Extractive summary: sentences quoted from the sources.