I believe everyone may be concerned about Google, which caused a sharp decline in gaming stocks at the beginning of the year.Project Genie” I’m still a bit impressed. The world model that can now be used in real-time generation of the game has a powerful new player. Yesterday ChinaTencentHunyuan team officially released and open sourced HY-World 2.0：A multi-modal 3D world model that takes input from text, single images, multi-view images or videos and directly generates real 3D assets (Mesh and 3D Gaussian Splatting) that can be used by game engines, rather than just outputting a pixel movie.
Key breakthrough: from “watching movies” to “building games”
The core proposition of HY-World 2.0 is a fundamental criticism of the existing film world model. Tencent clearly pointed out in its technical documents that models such as Google Genie 3, NVIDIA Cosmos and its previous generation HY-World 1.5 are essentially “generated pixel videos”. After the playback is completed, the generated world disappears and cannot be edited. HY-World 2.0 takes a completely different approach: direct output Realistic 3D assets(Mesh and 3DGS), these assets can be saved permanently and can be imported directly Blender, Unity, Unreal Engine, and NVIDIA Isaac Sim and other mainstream games and simulation engines.
We’re open-sourcing HY-World 2.0, a multimodal world model that generates, reconstructs, and simulates interactive *3D worlds* from text, images, and videos.
Outputs can be integrated into game engines and embodied simulation pipelines.
Key highlights:
🔹 One-click world… pic.twitter.com/OuKEm9krn4— Tencent HY (@TencentHunyuan) April 16, 2026
The differences between the two routes are clearly presented in a comparison table:
In short: the movie world model is “a movie that disappears after watching it”, while HY-World 2.0 is “a game world that stays forever after it is built.”
Four-stage technology pipeline
The world generation process of HY-World 2.0 is divided into four stages, which are interlocking:
- Panorama generation (HY-Pano 2.0): From text or image input, generate a 360° panorama as the basic frame of the world
- Trajectory planning (WorldNav): Plan the camera movement trajectory based on the panorama and determine the exploration path of the world.
- World expansion (WorldStereo 2.0): Follow a trajectory to expand a panorama into a 3D Gaussian Splatting world with depth and spatial consistency
- WorldMirror 2.0 + 3DGS Learning: Fusion of multi-view reconstruction results to output final high-quality 3DGS or Mesh format assets
In addition, HY-World 2.0 supportsWorld ReconstructionFunction: Input multiple photos or an ordinary video to instantly reconstruct the 3D scene of the digital twin: “Take a video and get a digital clone.”
WorldMirror 2.0: Single forward pass, outputting five types of information simultaneously
The core module of the entire system is WorldMirror 2.0, this is also the only sub-module that has been fully open source (code + model weight), and the number of parameters is approximately 1.2 billion (1.2B)。
The feature of WorldMirror 2.0 is that “single forward pass” can predict simultaneously:
- Dense Point Cloud
- Depth Map
- Surface Normal
- Camera parameters (Camera Intrinsics/Extrinsics)
- 3D Gaussian Splatting Properties
Supports flexible resolution inference from 50K to 500K pixels. On standard reconstruction benchmark tests such as 7-Scenes, NRGBD, and DTU, WorldMirror 2.0’s accuracy and completeness are both better than those of the first generation WorldMirror.
Performance Benchmark: Beyond SEVA, Gen3C
In the camera control evaluation of World Expansion (WorldStereo 2.0), HY-World 2.0 has obvious advantages over competing products:
- Rotation error (RotErr):0.492 (vs. SEVA’s 1.690, Gen3C’s 0.944)
- Displacement error (TransErr):0.968 (vs. SEVA’s 1.578, Gen3C’s 1.580)
- Visual quality (Q-Align):4.205 (vs. SEVA’s 3.232, Gen3C’s 3.353)
In the scene reconstruction F1 score of MipNeRF360, WorldStereo 2.0 reached 51.27, also leading the competition in the same period.
Interactive character exploration: “walk” in an AI-generated world
HY-World 2.0 is not just a static 3D asset output, but also supportsInteractive character exploration: Users can walk freely in AI-generated streets, buildings and landscapes from a first-person perspective or third-person character mode, with physical collision support.
Tencent has listed on its official website 3d.hunyuan.tencent.com Free trial is open (currently you have to wait in line due to heavy traffic).
Open source project: partially open, more on the way
As of the date of release, the open source status of HY-World 2.0 is as follows:
- ✅ Already open source：Technical Report
- ✅ Already open source: WorldMirror 2.0 inference code and model weights (Hugging Face can be downloaded)
- ⬜ Will be open source soon: Complete world generation inference code (WorldNav + WorldStereo 2.0 + World Composition)
- ⬜ Will be open source soon: Panorama generation model HY-Pano 2.0
- ⬜ Will be open source soon: The complete code and weights of the world expansion model WorldStereo 2.0
In other words, the current complete “one-click 3D world generation” process is not yet fully open source, but WorldMirror 2.0 (world reconstruction part) can be used directly. The complete build pipeline is expected to be released in subsequent batches. Model weights for WorldMirror 2.0 are currently available at Hugging Face Download, the complete world generation function is waiting for the subsequent open source process.
Source: KOCPC Chinese