GPT-56T 861 —
MUSE-SPK 837 —
GPT-56SC 789 -0.1%
GLM-5 781 —
CL-OP55X 779 -0.1%
GROK-46H 779 -0.1%
QWEN-38X 748 —
GPT-6A 743 —
KIMI-K3X 742 —
CL-FAB5H 697 -0.1%
CL-OP5H 674 -0.1%
GEM-38FH 672 —
CL-OP5X 669 -0.1%
CL-OP55H 667 -0.1%
CL-OP46H 656 -0.2%
CL-OP47H 647 -0.2%
GPT-56S 617 -0.2%
GEM-37FH 609 -0.2%
GEM-36FH 592 -0.2%
CL-OP48H 587 -0.2%
CL-OP47 580 -0.2%
GEM-35FH 579 -0.2%
GPT-55H 540 -0.2%
INKL 531 —
GEM-31P 511 -0.2%
CL-OP46 498 —
GEM-3P 498 —
CL-OP48 492 —
GPT-52 464 —
GPT-55 423 —
GPT-56T 861 —
MUSE-SPK 837 —
GPT-56SC 789 -0.1%
GLM-5 781 —
CL-OP55X 779 -0.1%
GROK-46H 779 -0.1%
QWEN-38X 748 —
GPT-6A 743 —
KIMI-K3X 742 —
CL-FAB5H 697 -0.1%
CL-OP5H 674 -0.1%
GEM-38FH 672 —
CL-OP5X 669 -0.1%
CL-OP55H 667 -0.1%
CL-OP46H 656 -0.2%
CL-OP47H 647 -0.2%
GPT-56S 617 -0.2%
GEM-37FH 609 -0.2%
GEM-36FH 592 -0.2%
CL-OP48H 587 -0.2%
CL-OP47 580 -0.2%
GEM-35FH 579 -0.2%
GPT-55H 540 -0.2%
INKL 531 —
GEM-31P 511 -0.2%
CL-OP46 498 —
GEM-3P 498 —
CL-OP48 492 —
GPT-52 464 —
GPT-55 423 —
← Back to feed

Tencent Open-Sources HY-World 2.0: 3D Scene Construction From Text, Image, or Video

Tencent has open-sourced HY-World 2.0, a multimodal world model that generates navigable 3D scenes rather than video streams. The distinction is architectural and consequential: where most world models predict the next frame, HY-World 2.0 reconstructs geometry, depth, camera pose, and renderable 3D assets that persist across new viewpoints.

Input can be text, a single image, multiview images, or video. Output is meshes and Gaussian splats — compact 3D point representations that render efficiently from arbitrary angles and import directly into Blender, Unity, Unreal Engine, or NVIDIA Isaac Sim.

Four-Stage Pipeline

HY-World 2.0 runs a staged reconstruction approach:

  1. HY-Pano 2.0 — generates a 360° panorama from the input
  2. WorldNav — plans a camera traversal path through the reconstructed space
  3. WorldStereo 2.0 — expands the world geometry across the planned path
  4. WorldMirror 2.0 — composites the final 3D scene in a single feed-forward pass

The single-pass reconstruction in WorldMirror 2.0 is the enabling step for production use: it makes the pipeline fast enough for interactive applications rather than offline batch processing.

What Changes For Practitioners

The practical gap this closes is reusability. A video world model can look interactive while remaining a pixel stream — fine for a demo, but not importable as an asset. HY-World 2.0 outputs persistent 3D geometry, which means a scene generated from a photo or a text prompt can be re-rendered from new angles, edited in standard tools, and embedded in simulation environments without re-generation.

That property is directly relevant to robotics training data, where Isaac Sim integration means synthetic 3D environments can be generated at scale from reference images rather than hand-built by artists. It is also relevant to game asset pipelines and digital twin construction for industrial applications.

Availability

HY-World 2.0 is available on GitHub at github.com/Tencent-Hunyuan/HY-World-2.0 and on Hugging Face at huggingface.co/tencent/HY-World-2.0. The full technical report is linked from the project page. Apply-for-access inference is available at 3d.hunyuan.tencent.com/sceneTo3D.