HY-World 2.0: Best AI World Models for 3D Teams in 2026

What Is HY-World 2.0?

HY-World 2.0 is Tencent Hunyuan's open-source AI World Model for reconstructing, generating, and simulating editable 3D scenes, and it is one of the best AI World Models tools for 3D researchers, game developers, robotics teams, and simulation engineers. The repo says WorldMirror 2.0 can infer depth, normals, camera parameters, point clouds, and 3D Gaussian Splatting in a single forward pass across 50K–500K pixel inputs, with the first public technical report and partial code released on Apr 16, 2026.

The important part is not just output quality. HY-World 2.0 produces persistent 3D assets that can be imported into Blender, Unity, Unreal Engine, and Isaac Sim, which makes it materially different from video-only world models that only emit frames.

Quick Overview

Who Should Use HY-World 2.0?

• 3D graphics teams building scene generation pipelines that need editable outputs instead of flat video. HY-World 2.0 emits meshes and 3DGS, so the result can move downstream into DCC tools and real-time engines.
• Robotics and simulation engineers who care about navigation, collision, and persistent geometry. The repo explicitly positions the model for physical interaction, not just visual playback.
• Game developers prototyping explorable environments from a prompt, a concept image, or a capture sequence. HY-World 2.0 is useful when you need a seed world that can be iterated in Unity or Unreal rather than hand-modeled from scratch.
• Applied AI researchers evaluating world models, 3D reconstruction, and feed-forward scene understanding. The architecture combines generation, reconstruction, and composition into one research stack.

Not ideal for:

• Teams that only need a short cinematic clip and do not care about editable geometry.
• Shipping products that require a fully stable API today, because parts of the full generation stack are still marked as coming soon.
• Users without access to a GPU environment, since 3D reconstruction and splatting workflows are not lightweight CPU-only jobs.

Key Features of HY-World 2.0

• Text-to-3D world generation — HY-World 2.0 can start from a text prompt and synthesize navigable 3D scenes rather than a transient video stream. The generation path is staged through panorama creation, trajectory planning, world expansion, and final composition.
• Single-image world bootstrapping — A single concept image can seed a 3D world. That is useful for concept art, level design, and rapid environment prototyping when you have one reference but not a full capture sequence.
• Unified feed-forward reconstruction — WorldMirror 2.0 predicts depth, surface normals, camera parameters, 3D point clouds, and 3DGS attributes in one pass. That design reduces the need for a separate stack of monocular depth, pose estimation, and splat fitting tools.
• Multi-modal input support — The system accepts text, single-view images, multi-view images, and videos. That makes HY-World 2.0 practical for both synthetic generation and real-world capture reconstruction.
• Editable 3D outputs — The model produces meshes, point clouds, and Gaussian splats that can be edited, lit, and re-exported. This is the core reason HY-World 2.0 is more useful than a video world model for production workflows.
• Engine-friendly integration — The repo says outputs can be imported into Blender, Unity, Unreal Engine, and Isaac Sim. That matters if the next step is simulation, interaction, or offline content editing rather than passive viewing.
• Interactive exploration — HY-World 2.0 supports first-person navigation and third-person character mode with physics-based collision. That turns generated scenes into something closer to a playable environment than a rendered demo.

HY-World 2.0 vs Alternatives

Pick Genie 3 when you want a video-first experience and do not need editable geometry. Pick Cosmos when you are comparing large-scale simulation research systems and care more about model family breadth than asset export.

Pick HY-World 1.5 only if you need to study the earlier Tencent pipeline or compare against prior art. If you want surrounding tooling for pipeline tracing or agent-driven evaluation, pair the workflow with OpenTrace or OpenSwarm. For broader exploration, browse all AI tools or 3D tools.

How HY-World 2.0 Works

HY-World 2.0 uses a two-track architecture: one branch for world generation and one branch for world reconstruction. The generation branch is explicitly staged, which is a sane design choice because panorama synthesis, camera path planning, spatial expansion, and final scene composition are different problems with different failure modes.

The reconstruction branch centers on WorldMirror 2.0, a feed-forward model that maps multi-view images or casual video into geometry and splat attributes in one pass. That means the model is not just hallucinating appearance; it is estimating scene structure, camera geometry, and renderable 3D representations that survive outside the inference session.

git clone https://github.com/Tencent-Hunyuan/HY-World-2.0.git
cd HY-World-2.0
pip install -r requirements.txt
python tools/worldmirror_infer.py --input samples/demo.mp4 --output outputs/demo

That example clones the repo, installs dependencies, and runs a reconstruction pass on a sample video. Expect outputs such as depth maps, normals, camera estimates, point clouds, and 3DGS artifacts, then export them into your downstream renderer or editor.

Pros and Cons of HY-World 2.0

Pros:

• Outputs real 3D assets instead of only producing frames, which makes the results usable in game engines and simulation stacks.
• Multi-modal input coverage covers text, single images, multi-view captures, and video, so the same project can handle synthetic and real-world sources.
• Feed-forward reconstruction reduces the need to chain separate depth, pose, and splatting systems together.
• Engine interoperability with Blender, Unity, Unreal Engine, and Isaac Sim is a strong practical advantage for content and simulation teams.
• Interactive navigation with first-person and third-person modes makes validation easier than inspecting a flat render set.

Cons:

• Full generation code is not fully released yet, so the most interesting branch is still partial as of Apr 16, 2026.
• 3D workflows are GPU-hungry, so this is not a casual laptop tool.
• The repo is research-first, which means setup, weights, and inference paths may change between releases.
• World-model quality depends on source inputs, so poor videos or ambiguous images still produce messy geometry.
• Licensing and deployment details are not fully visible from the page text, so product teams need to verify usage terms before shipping.

Getting Started with HY-World 2.0

git clone https://github.com/Tencent-Hunyuan/HY-World-2.0.git
cd HY-World-2.0
conda create -n hyworld python=3.10 -y
conda activate hyworld
pip install -r requirements.txt
python tools/worldmirror_infer.py --input samples/demo.mp4 --output outputs/demo

After the first run, you should expect the repo to download or locate model weights, then emit 3D reconstruction artifacts into the output directory. If you are using the public model releases, you will likely need to pull weights from Hugging Face or ModelScope and point the scripts at the local checkpoint path.

Before running a real workload, verify CUDA availability, confirm the expected Python version, and allocate enough VRAM for the chosen resolution. The repo calls out flexible-resolution inference, so your first tuning pass should be input size and memory budgeting rather than prompt engineering.

Verdict

HY-World 2.0 is the strongest option for teams that need editable 3D worlds when they can tolerate a research-grade release cycle. Its biggest strength is turning multimodal inputs into persistent geometry, and its main caveat is that the full generation stack is still incomplete. If your workflow ends in Blender, Unity, Unreal, or Isaac Sim, HY-World 2.0 is worth serious evaluation.