HY-World 2.0 Complete Guide

HY-World 2.0 shipped on 15 April 2026 as an open multimodal world model stack. It accepts text, a single image, multi-view images or video, and aims at 3D representations: Gaussian splats, meshes and reconstruction buffers, not a flat MP4.

The four-stage generation path

For text or a single image the paper describes panorama generation with HY-Pano 2.0, trajectory planning with WorldNav, world expansion with WorldStereo 2.0, and composition with WorldMirror 2.0. WorldLens is the viewer: engine-agnostic rendering, IBL lighting, collision hints and character support.

Reconstruction is the other half. WorldMirror 2.0 is the feed-forward model most public write-ups actually run first. Sister stages were marked coming soon in early repo notes, so check the GitHub README before you assume the full pipeline is one checkpoint.

How it compares

Tencent positions HY-World 2.0 against closed systems such as World Labs Marble: open weights, multi-modal input, navigable output. It is still a world representation, not a shipped game. If you need a playable loop this week, plan here and build in OpenGame. If you need a splat or mesh to inspect, HY-World is the closer tool.