HunyuanWorld

model
Tencent
Open-source world generation model series. World 1.0 integrates panoramic visual generation and 3D reconstruction from text/image inputs. HY-World 1.5 (WorldPlay) adds real-time interactive streaming at 24fps with Context Forcing distillation and WorldCompass RL post-training.
Paper (arXiv)GitHub (World 1.0)GitHub (HY-WorldPlay / World 1.5)HuggingFace (HY-WorldPlay)

Outputs 3

HunyuanWorld 1.0

model

First open-source world model supporting immersive roaming, interaction, and simulation from text and image inputs.

Paper (arXiv)GitHub
arXiv HTML

HY-World 1.5 (WorldPlay)

model

Real-time interactive world model with streaming video diffusion at 24fps. Introduces Context Forcing distillation and WorldCompass RL post-training framework.

GitHubHuggingFace

HY-World 2.0

model

Multi-modal world model that reconstructs, generates, and simulates 3D worlds from text, images, or video. Produces editable 3D Gaussian Splatting scenes directly usable in Unreal/Unity/Blender. Four-stage pipeline: HY-Pano 2.0 (panorama generation), WorldStereo 2.0 (keyframe view generation), WorldMirror 2.0 (feed-forward 3D prediction), and WorldLens (engine-agnostic rendering with IBL lighting). Performance comparable to closed-source Marble. Open weights and code.

Paper (arXiv)GitHub
arXiv HTML
3dgenerationembodiedopen-weight