Key Features

• Lyra 1.0: feed-forward 3D and 4D scene generation from a single image or video via diffusion self-distillation
• Lyra 2.0: explorable generative 3D worlds with long-horizon, 3D-consistent generation
• Gaussian splatting output format enables real-time rendering of generated scenes
• Pre-trained weights available on HuggingFace for both Lyra 1.0 and 2.0
• Apache 2.0 code license permits commercial experimentation and modification
• arXiv papers included for full technical documentation of both model versions
• Modular repository structure with dedicated subdirectories for Lyra-1 and Lyra-2
• Video diffusion model self-distillation enables 3D learning without multi-view capture