Lyra 2.0 is an Apache 2.0 licensed video diffusion and generative world model developed by NVIDIA, designed to generate large-scale 3D scenes and explorable three-dimensional environments from single input images. Released in 2026, the system represents an advancement in generative 3D technology, enabling the creation of persistent, interactive worlds through innovative approaches to geometry tracking and self-augmented training methodology.¹⁾-computer|ThursdAI (2026]]))

Lyra 2.0 addresses significant challenges in 3D scene generation and video synthesis: maintaining spatial consistency, geometric coherence, and temporal stability across video sequences. Rather than treating video generation as a purely 2D problem, the system leverages per-frame 3D geometry as a foundational mechanism for organizing and retrieving historical information about the scene being generated. This geometric grounding enables the model to maintain 3D spatial awareness throughout the generation process, resulting in worlds that remain explorable and internally consistent from multiple viewpoints ²⁾.

The open-source licensing under Apache 2.0 facilitates broad adoption and community-driven development within the generative modeling research community.

The core innovations in Lyra 2.0 address two critical failure modes common in video generation and 3D synthesis: spatial forgetting and temporal drifting. Spatial forgetting refers to the degradation or inconsistency of spatial geometry across generated frames, while temporal drifting describes the accumulation of visual inconsistencies over extended generation sequences.

Per-frame 3D geometry retrieval serves as a foundational mechanism to maintain spatial consistency. As the model generates successive frames of video, it maintains explicit 3D geometric representations that allow it to query and reference previously generated spatial information. This approach contrasts with purely sequential attention mechanisms, instead creating a structured 3D space that anchors the generation process. Objects, surfaces, and spatial relationships maintain coherence throughout the generated sequence ³⁾.

Self-augmented training represents another key technical component. Rather than relying solely on paired training data, the system employs self-generated content to augment its training distribution. The model trains on progressively corrupted or degraded versions of its own outputs, learning recovery mechanisms and maintaining quality even when encountering partially corrupted or noisy intermediate representations. This methodology enables better generalization across diverse 3D scene types and input conditions, and allows the system to improve consistency and quality through iterative refinement of its own outputs ⁴⁾.

Lyra 2.0's capability to generate explorable 3D worlds from single images positions it for applications across multiple domains:

* 3D scene generation from photographs: Users can input a single image and generate extended, explorable 3D environments with consistent geometry and appearance * Interactive world exploration: Generated worlds can be explored from multiple viewpoints, enabling virtual camera movement and perspective shifts * Entertainment and game development: The system accelerates environment creation by automatically generating complete 3D scenes from concept art or reference images * Architectural visualization: The technology enables rapid generation of immersive walkthroughs from static architectural plans or photographs * Content creation workflows: The model accelerates 3D asset creation for games, films, and architectural visualization * Scientific visualization and simulation: The system facilitates the creation of complex 3D environments for research and training applications ⁵⁾.

• Tencent HYWorld 2.0 vs NVIDIA Lyra 2.0
• WorldMirror 2.0
• Video Diffusion Models for 3D World Generation