Tencent released HY-World 2.0, an open-source framework that converts a text prompt, a single photo, or a video into a navigable 3D world. Unlike video world models like Genie 3 that produce flat clips, HY-World 2.0 outputs real 3D assets: meshes and 3D Gaussian Splats (a technique that represents scenes as collections of colored 3D blobs for fast, high-quality rendering) that import directly into Unity, Unreal Engine, or Blender. The pipeline runs four stages: panorama generation, trajectory planning with 3D scene understanding, world expansion with memory modules that enforce geometric consistency across views, and final composition via WorldMirror 2.0, a 1.2B-parameter reconstruction model. It ships with WorldLens, a rendering platform with collision detection and character support. On benchmarks, it matches the closed-source model Marble and leads all open-source alternatives.
A game designer can type “medieval village at sunset” and get a playable 3D environment in minutes instead of weeks of manual modeling. Generated assets are engine-compatible, slotting into existing workflows without conversion. The memory architecture is what makes this practical: earlier open-source attempts at 3D world generation drifted into incoherence as scenes grew. HY-World 2.0 maintains consistency across views, producing worlds stable enough to actually use.
This signals a move from video world models to 3D-native ones. Video models generate pixels you watch. 3D models generate geometry you interact with: zero-error camera control, real physics, and permanent assets that render in real time on consumer hardware. Open-sourcing the full pipeline lets the community improve each stage independently.
Read More: World Labs raised $1B for similar spatial intelligence work, but kept its models closed.
Sources:
- HY-World 2.0 Paper (arXiv)
- HY-World 2.0 GitHub Repository
- HY-World 2.0 Project Page
- HuggingFace Daily Papers
Disclaimer: For information only. Accuracy or completeness not guaranteed. Illegal use prohibited. Not professional advice or solicitation. Read more: /terms-of-service
Reuse
Citation
@misc{kabui2026,
author = {{Kabui, Charles}},
title = {HY-World 2.0: {Tencent’s} {Open-Source} {Model} {Turns}
{Text} or a {Photo} {Into} a {Walkable} {3D} {World}},
date = {2026-04-20},
url = {https://toknow.ai/posts/tencent-hy-world-2-open-source-3d-world-model-text-image/},
langid = {en-GB}
}