Agent skill · creative production · aradotso
hy-world-2-0-3d-world-model
Expert skill for using HY-World 2.0, Tencent's multi-modal world model for reconstructing, generating, and simulating 3D worlds from text, images, and video.
Why this skill is useful
Adds a complete pipeline for 3D world reconstruction and generation, including specific model usage and installation instructions that are not widely known.
What it needs
Requires cuda, flash-attn, python, pytorch installed locally. About 5k tokens when loaded. Last updated 2026-07-09. 70 stars on the source repository.
What this skill does
HY-World 2.0 — 3D World Model Skill Skill by ara.so — Daily 2026 Skills collection. HY-World 2.0 is a multi-modal world model by Tencent Hunyuan that reconstructs, generates, and simulates 3D worlds. It accepts text, single-view images, multi-view images, and videos as input and produces 3D representations (meshes, 3D Gaussian Splattings, point clouds). Two core capabilities: World Reconstruction (multi-view images / video → 3D): Powered by WorldMirror 2.0, a ~1.2B feed-forward model predicting depth, surface normals, camera parameters, 3D point clouds, and 3DGS attributes in a single forward pass. World Generation (text / single image → 3D world): Four-stage pipeline — Panorama Generation (HY-Pano 2.0) → Trajectory Planning (WorldNav) → World Expansion (WorldStereo 2.0) → World Composition (WorldMirror 2.0 + 3DGS). --- Installation Requirements Python 3.10 CUDA 12.4 (recommended) PyTorch 2.4.0 Model Weights Model weights are automatically downloaded from Hugging Face on first run. Alternatively, download manually: Model HuggingFace --- --- WorldMirror 2.0 tencent/HY-World-2.0 → HY-WorldMirror-2.0 WorldMirror 1.0 (legacy) tencent/HunyuanWorld-Mirror To pre-download: --- Core API — WorldMirror 2.0 (World Reconstruction) Basic Usage With Prior Injection (Camera & Depth) Provide known camera parameters or depth priors to improve accuracy: Camera JSON Format The priorcamera.json format expected by the pipeline: Result Object The pipeline returns a result object with the following attributes: --- Gradio App — WorldMirror 2.0 Launch an interactive web UI for 3D reconstruction: Access at http://localhost:7860 by default. --- Common Patterns Pattern 1: Reconstruct from a Video Extract frames from a video, then run reconstruction: Pattern 2: Flexible Resolution Inference WorldMirror 2.0 supports 50K–500K pixel resolution. …
How to use it
Reference it in AdaL, Claude Code, Cursor or any coding agent — nothing to install:
@skills aradotso/hy-world-2-0-3d-world-model