Latent Diffusion Video Synthesis Engine

via “text-to-video generation with dit-based diffusion”

Official repository for LTX-Video

Unique: First DiT-based video generation model optimized for real-time inference, generating 30 FPS videos faster than playback speed through causal video autoencoder latent-space diffusion with rectified flow scheduling, enabling sub-second generation times vs. minutes for competing approaches

vs others: Generates videos 10-100x faster than Runway, Pika, or Stable Video Diffusion while maintaining comparable quality through architectural innovations in causal attention and latent-space diffusion rather than pixel-space generation