tencent-hunyuan/hy-world-2.0

HY-World 2.0 is an AI system that converts text descriptions, photos, or videos into real, interactive 3D environments, not just video clips, but actual 3D geometry you can explore and edit. The core idea is that most AI "world models" only generate video, which disappears once it plays. HY-World 2.0 instead produces persistent 3D assets (meshes and what's called 3D Gaussian Splatting, a technique for representing 3D scenes) that can be imported directly into tools like game engines. The system works in four stages: first it generates a 360-degree panoramic image from your text or photo, then it plans a camera path through the scene, then it expands the scene into a full 3D world, and finally it assembles everything into a navigable 3D asset. For reconstruction from existing photos or videos, a component called WorldMirror 2.0 estimates depth, surface normals, and camera positions in a single pass to build a 3D digital twin. You'd use this if you want to generate a 3D environment from a description or image, for game development prototyping, architectural visualization, or AI research into 3D scene generation. It supports walking through the generated world in first-person or third-person view with physics-based collision. The project is Python-based, open-source, and model weights are available on HuggingFace. The full README is longer than what was provided.