Tencent-Hunyuan/HY-World-2.0

What it solves

HY-World 2.0 addresses the limitations of video-based world models, which produce non-editable pixel videos that vanish after playback. Instead, it generates persistent, editable 3D assets (meshes, Gaussian Splattings, and point clouds) that can be integrated into game engines like Unity, Unreal Engine, and Blender, enabling unlimited playable duration and native 3D consistency.

How it works

The framework provides two primary capabilities:

1. World Generation: Converts text or single images into 3D worlds via a four-stage pipeline:
   • Panorama Generation (HY-Pano 2.0) creates a 360° view.
   • Trajectory Planning (WorldNav) plans camera movement with obstacle awareness.
   • World Expansion (WorldStereo 2.0) generates keyframes for the world.
   • World Composition (WorldMirror 2.0 + 3DGS learning) assembles the final 3D representation.
2. World Reconstruction: Uses WorldMirror 2.0, a feed-forward model that predicts depth, surface normals, camera parameters, 3D point clouds, and 3DGS attributes from multi-view images or videos in a single pass.

Who it’s for

It is designed for researchers and developers in 3D scene generation, digital twin creation, and game development who need navigable, physics-compliant 3D environments rather than simple video clips.

Highlights

• Real 3D Assets: Produces 3DGS and meshes instead of flat videos, allowing for real-time rendering on consumer GPUs.
• Instant Reconstruction: Creates digital twins from casual videos or photos using a unified feed-forward model.
• Interactive Exploration: Supports first-person and third-person navigation with physics-based collision within generated worlds.
• Engine Compatibility: Assets are directly importable into Blender, Unity, Unreal Engine, and Isaac Sim.