We’re excited to announce that LightX2V is expanding from video generation model inference to world modeling capabilities. LightX2V now supports two cutting-edge world models: Matrix Game 3.0 for real-time interactive video generation and HY-WorldMirror 2.0 for 3D world reconstruction. These integrations bring significant performance improvements and new capabilities to our unified inference platform.

Matrix Game 3.0: Real-Time Interactive World Modeling

Matrix Game 3.0 by Skywork AI is a memory-augmented interactive world model that achieves up to 40 FPS real-time generation at 720p resolution. It’s the first industrial-scale deployable world model for interactive applications, supporting mouse and keyboard inputs with long-term memory consistency over minute-long sequences.

Performance Improvements in LightX2V

Our integration (PR #989) delivers significant performance improvements over the official implementation:

Matrix Game 3.0 Performance Performance comparison between LightX2V implementation and Matrix-Game official

Matrix Game 3.0 real-time interactive generation demo

Key integration benefits:

  • Enhanced memory efficiency: Optimized attention mechanisms for long sequences
  • Improved hardware utilization: Better performance on A800 GPUs with FlashAttention2
  • Streamlined deployment: Unified inference APIs with existing LightX2V workflows

HY-WorldMirror 2.0: 3D World Reconstruction

HY-WorldMirror 2.0 by Tencent’s Hunyuan team generates actual 3D assets instead of temporary video files. The ~1.2B parameter model supports both world generation from text/images and world reconstruction from multi-view inputs.

The model creates navigable 3D worlds through a 4-stage pipeline from panorama generation to final 3D scene composition for world generation, while its reconstruction capability predicts depth, surface normals, camera parameters, 3D point clouds, and 3DGS attributes in a single forward pass. It achieves state-of-the-art results with 0.012 accuracy and 0.016 completeness on the 7-Scenes dataset, along with 35% improvement in rotation accuracy compared to version 1.0.

HY-WorldMirror Performance Performance acceleration comparison between LightX2V implementation and baseline

HY-WorldMirror 2.0 3D world reconstruction demo

Our integration PR #1022) delivers significant performance acceleration with multi-GPU scaling through sequence-parallel inference for large scenes, advanced FP8 quantization support with real-time calibration, Intel XPU acceleration through SYCL/ESIMD kernels, and memory optimization via CPU offloading and lazy loading for efficient inference.

Getting Started

Both models are available in the latest LightX2V release. Check our documentation for installation and usage examples.

Matrix Game 3.0: Perfect for real-time interactive applications requiring responsive world modeling.

HY-WorldMirror 2.0: Ideal for 3D content creation and applications requiring persistent 3D assets.

Looking Forward

These integrations demonstrate LightX2V’s commitment to bringing cutting-edge world modeling technologies to developers through optimized, production-ready implementations. We continue to push the boundaries of what’s possible in AI-powered world creation.

LightX2V is an open-source unified framework for X-to-Video generation. Follow us on GitHub for the latest updates.