Hunyuan3D World Model 2.0 - Tencent Hunyuan's Open-Source Multimodal World Model

What is Hunyuan3D World Model 2.0

Hunyuan3D World Model 2.0 (HY-World 2.0) is an open-source multimodal world model launched by Tencent Hunyuan. It supports generating roaming-capable 3D Gaussian Splatting (3DGS) scenes from text, a single image, multiple views, or video. The model adopts a four-stage architecture: panorama generation → trajectory planning → world expansion → 3D reconstruction, achieving the unification of "text/image-to-3D world" and "video reconstruction of 3D world." Scenes generated by Hunyuan3D World Model 2.0 support physical collision and character exploration, and can be exported to Unity/UE engines, with performance benchmarked against closed-source commercial products.Marble。

Main Features of Hunyuan3D World Model 2.0

• World Generation: Generates 360° roaming-capable 3DGS/Mesh immersive scenes based on text or a single image.
• World Reconstruction: Reconstructs high-fidelity 3D digital twin spaces from multi-view image or video input.
• Panorama Generation: Converts images from any viewpoint or text into 360° panoramas (HY-Pano 2.0).
• Intelligent Trajectory Planning: Parses scene semantics and plans exploration paths, avoiding unreasonable behaviors such as walking through walls (WorldNav).
• Character Adventure Mode: Supports controlling a character to walk and explore freely in generated scenes.
• Multi-Format Export: Supports multiple formats such as 3DGS, Mesh, point cloud, and video, as well as integration with mainstream engines.

Technical Principles of Hunyuan3D World Model 2.0

• Panorama Generation (HY-Pano 2.0): Adopts an end-to-end implicit learning scheme, using a multimodal Diffusion Transformer (MMDiT) to autonomously learn the spatial mapping from perspective views to 360° Equirectangular Projection (ERP), without requiring camera metadata. Introduces Circle Padding and pixel blending techniques to solve the ERP boundary discontinuity problem, combined with hybrid training on real panoramas + UE synthetic data.
• Trajectory Planning (WorldNav): Understands spatial structure through geometric and semantic scene parsing (point cloud, Mesh, semantic segmentation, NavMesh), intelligently plans roaming trajectories that maximize information coverage, ensures paths are natural and avoid obstacles, and supports diverse trajectory types such as regular, surround, reconstruction-aware, roaming, and aerial.
• World Expansion (WorldStereo 2.0): Based on keyframe generation rather than continuous video generation, introduces global geometric memory and spatial stereo memory mechanisms to ensure multi-view consistency. Improves efficiency through model distillation, enabling scene expansion and detail completion along planned paths.
• World Reconstruction (WorldMirror 2.0): Adopts a feed-forward 3D prediction model, optimizing geometric accuracy through normalized positional encoding, explicit normal supervision, and depth mask prediction. Combines 3D Gaussian Splatting (3DGS) for final scene fusion and optimization, supporting robust reconstruction from generated views and real input.

How to Use Hunyuan3D World Model 2.0

• Visit the Platform: Visit theTencent Hunyuan 3Dofficial website and log in to your account..
• Select Input Method: Upload a single image (any viewpoint supported) or enter a text description (e.g., "cyberpunk-style street with flickering neon lights")
• Parameter Configuration: Select parameters such as generation mode (world generation/world reconstruction), scene style, and roaming range.
• Generation and Exploration: After clicking Generate, the system automatically executes a four-stage process (panorama → trajectory → extension → reconstruction).
• Export and Usage: Supports downloading 3DGS/Mesh files or directly generating asset packages importable into the Unity/UE engines.

Key Information and Usage Requirements of Hunyuan 3D World Model 2.0

• Development Team: Developed and launched by Tencent Hunyuan.
• Open Source Status: Model weights, training code, and the complete technical report have been fully open-sourced and can be freely downloaded for research.
• Input Modalities: Supports four input methods: text prompts, a single image, multi-view images, and video streams.
• Output Formats: Can export 3D Gaussian Splatting (3DGS), mesh models (Mesh), point cloud data, 360° panoramas, and video sequences.
• Engine Support: Generated assets are natively compatible with Unity and Unreal Engine (UE), and can be directly imported for further editing.
• Hardware Requirements: Local deployment requires an NVIDIA GPU; an environment with ≥16GB of VRAM is recommended to ensure generation efficiency.

Core Advantages of Hunyuan 3D World Model 2.0

• Unified Generation and Reconstruction: The first open-source framework to simultaneously support both paradigms: "creating from nothing" (generation) and "from real to virtual" (reconstruction).
• Physically Interactable: Generated scenes possess spatial logic and physical collision detection, supporting character-controlled exploration.
• No Camera Parameters Required: The panorama generation stage requires no metadata such as focal length or FOV, adapting to real-world scene images.
• High-Quality Boundary Handling: 360° panoramas connect seamlessly, without the left-right boundary seam issues of traditional ERP images.
• Memory Consistency: The world extension stage introduces a memory mechanism to ensure visual coherence during long-trajectory roaming.
• Industrial-Grade Compatibility: Natively supports mainstream game engines and can be directly used for game development and robot simulation.

Project Addresses of Hunyuan 3D World Model 2.0

• Project Official Website：https://3d-models.hunyuan.tencent.com/world/
• GitHub Repository：https://github.com/Tencent-Hunyuan/HY-World-2.0
• HuggingFace Model Library：https://huggingface.co/tencent/HY-World-2.0
• Technical Paper：https://3d-models.hunyuan.tencent.com/world/world2_0/HY_World_2_0.pdf

Comparison of Hunyuan 3D World Model 2.0 with Similar Competitors

Application scenarios of Hunyuan3D World Model 2.0

• Game development: Developers can use the model to quickly generate interactive 3D scene prototypes with physical collision, significantly reducing the time and technical barriers of traditional level design.
• Virtual reality applications: The system can build high-fidelity immersive VR/AR spaces, supporting end users to walk freely and explore in real time within the generated environment.
• Digital twin construction: Users only need to provide a small number of photos or short videos to automatically reconstruct a high-precision 3D digital copy of the real physical world.
• Robot simulation training: Provides a physically consistent and interactive virtual training environment for embodied intelligence and autonomous driving systems, supporting safe algorithm validation.
• Film and television virtual production: The model can generate 360° panoramic virtual background assets for movie previsualization, virtual studio shooting, and post-production visual effects compositing.