Lyra SDG: 59,031 Synthetic Multi-View Images for 3D Scene Reconstruction

NVIDIA researchers released this synthetic multi-view 3D and 4D dataset in September 2025, containing 59,031 images generated via the GEN3C framework. It serves as the training foundation for Lyra, a generative model utilizing video diffusion self-distillation for scene reconstruction.

Use Cases

• Training image-to-3d generative models using the 59,031 multi-view images
• Refining video diffusion models through self-distillation for spatial intelligence
• Benchmarking Gaussian Splatting algorithms on synthetic 4D scene data

Strengths

• Contains 59,031 high-quality synthetic images
• Supports both 3D and 4D scene reconstruction tasks
• Permissive CC-BY-4.0 licensing

Limitations

• Synthetic nature may result in a domain gap when applied to real-world camera footage
• Lack of detailed column-level documentation in the initial release

Provenance

Source

NVIDIA Research

Collection Method

Synthetic data generation using the GEN3C framework

Freshness

Last updated September 2025.