Hunyuan World 2.0 (3D World Model)

Deploy Tencent HY-World 2.0 — the first open-source SOTA 3D world model — on Clore.ai GPUs for text/image/video-to-3D scene generation

Released April 15, 2026 — Tencent Hunyuan dropped HY-World 2.0, the first fully open-source SOTA 3D world model. This guide covers WorldMirror 2.0 (the shipped ~1.2B parameter reconstruction component). Sister models HY-Pano 2.0 and WorldStereo 2.0 are flagged "coming soon" in the official repo — see the Roadmap below.

HY-World 2.0 is Tencent's multi-modal world-model framework for reconstructing, generating, and simulating full 3D scenes. Unlike single-object mesh generators, HY-World ingests text, single or multi-view images, or video and emits editable world representations — meshes, 3D Gaussian Splats, point clouds, depth maps, surface normals, and recovered camera parameters — ready to drop into Unity, Unreal, or Blender.

The first public weights cover WorldMirror 2.0 (~1.2B params, BF16) — the reconstruction half of the stack. It runs in ~12–24 GB of VRAM on a single GPU and supports flexible resolution from 50K to 500K pixels, plus FSDP multi-GPU sharding for larger workloads. A Python API (diffusers-style), CLI via torchrun, and a Gradio demo ship out of the box. A ComfyUI node is not official yet — community ports only.

All examples in this guide run on GPU servers rented through the CLORE.AI Marketplace.

Key Specs

Property	Value
Component	WorldMirror 2.0 (shipped); HY-Pano 2.0 + WorldStereo 2.0 coming soon
Parameters	~1.2B (BF16)
Input modalities	Text · single-view image · multi-view images · video
Output	Mesh · 3D Gaussian Splat · point cloud · depth · normals · camera params
VRAM	~12–24 GB single GPU; FSDP for multi-GPU
Resolution range	50K – 500K pixels (flex-res)
License	tencent-hy-world-2.0-community (custom — see below)
Release	2026-04-15

License caveat: HY-World 2.0 ships under a custom community license (License.txt at repo root), not Apache 2.0 or MIT. Commercial use terms differ from Tencent's Hunyuan3D 2.1. Read the full license before shipping anything built on it.

Why HY-World 2.0?

• First open-source SOTA world model — no closed competitors in this category

• Full scene output, not just meshes — Gaussian Splats + geometry + camera in one pass

• Multi-modal inputs — same pipeline handles text, images, and video

• FSDP-ready — scale across 2–8 GPUs for high-res or batched inference

• Game-engine ready — outputs drop straight into Unity, Unreal, and Blender

---

Requirements

Component	Minimum	Recommended
GPU VRAM	16 GB (RTX 4080 / 3090)	24–80 GB (RTX 4090 / A100 / H100)
System RAM	32 GB	64–128 GB
Disk	80 GB	200 GB
CUDA	12.1	12.4+
Python	3.10	3.10
PyTorch	2.4.0	2.4.0+

Multi-GPU mode requires ≥ 1 input image per GPU. For a single reference image, stick with one GPU and let FSDP kick in only for batched or high-resolution jobs.

---

Option A — Quickstart with Docker + torchrun

A minimal docker-compose.yml for a Clore.ai container (official Tencent image is not yet published — this uses the PyTorch base and runs the repo install inside):

version: "3.8"
services:
  hyworld2:
    image: pytorch/pytorch:2.4.0-cuda12.4-cudnn9-devel
    ports:
      - "7860:7860"
    volumes:
      - ./workspace:/workspace
      - hf_cache:/root/.cache/huggingface
    working_dir: /workspace
    command: >
      bash -c "
        git clone https://github.com/Tencent-Hunyuan/HY-World-2.0 &&
        cd HY-World-2.0 &&
        pip install -r requirements.txt &&
        pip install flash-attn --no-build-isolation &&
        python -m hyworld2.worldrecon.gradio_app
      "
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]
    shm_size: "16gb"

volumes:
  hf_cache:

Run a multi-GPU reconstruction job with FSDP and BF16:

torchrun --nproc_per_node=2 -m hyworld2.worldrecon.pipeline \
    --input_path /workspace/input_images \
    --use_fsdp --enable_bf16

---

Option B — Manual Python API

# Clone and install
git clone https://github.com/Tencent-Hunyuan/HY-World-2.0
cd HY-World-2.0
conda create -n hyworld2 python=3.10 -y
conda activate hyworld2
pip install torch==2.4.0 torchvision==0.19.0 --index-url https://download.pytorch.org/whl/cu124
pip install -r requirements.txt
pip install flash-attn --no-build-isolation

from hyworld2.worldrecon.pipeline import WorldMirrorPipeline

# Loads ~1.2B BF16 weights from HF (tencent/HY-World-2.0)
pipeline = WorldMirrorPipeline.from_pretrained('tencent/HY-World-2.0')

# Reconstruct a 3D scene from a folder of multi-view images
result = pipeline('path/to/images')

# Optional: inject prior camera + depth for tighter reconstruction
result = pipeline(
    'path/to/images',
    prior_cam_path='path/to/prior_camera.json',
    prior_depth_path='path/to/prior_depth/',
)

Launch the Gradio demo on port 7860:

python -m hyworld2.worldrecon.gradio_app

For multi-GPU Gradio with FSDP:

torchrun --nproc_per_node=2 -m hyworld2.worldrecon.gradio_app \
    --use_fsdp --enable_bf16

---

Clore.ai GPU Recommendations

Workload	GPU	VRAM	Why	Clore.ai Cost
Single image → scene, dev/preview	RTX 4090	24 GB	BF16 fits comfortably, fast iteration	~$0.5–2/day
Multi-view video reconstruction	A100 40 GB	40 GB	Handles 200K+ px frames without OOM	~$3–5/day
High-res batched (production)	A100 80 GB	80 GB	Full 500K px flex-res, big batches	~$5–8/day
FSDP multi-GPU / research	2–4× H100	160–320 GB	Sharded training-scale workloads	~$15–40/day

Sweet spot on Clore.ai: a single RTX 4090 at ~$0.5–2/day handles everyday WorldMirror inference. Step up to an A100 only when you need >200K-pixel reconstructions or long video inputs.

---

Use Cases

• Game development — turn concept art into rough 3D environments for blockout and greybox

• AR/VR content — generate Gaussian Splat scenes playable in Unity/Unreal with near-photographic fidelity

• Film and animation previs — reconstruct sets from on-location photos for virtual cinematography

• Architectural visualization — convert reference shots or text briefs into editable 3D walkthroughs

• Robotics + simulation — synthesize 3D training environments from sparse real-world footage

---

Roadmap

Tencent has listed the following as "coming soon" in the official repo:

• HY-Pano 2.0 — 360° panorama generation (interim: HunyuanWorld 1.0)

• WorldStereo 2.0 — world expansion / novel-view synthesis (interim: original WorldStereo)

• WorldNav — trajectory planning for scene traversal

• Full world-generation pipeline code — the text/image → full world entry point

WorldMirror 2.0 (reconstruction) is the only component with public weights today. Keep an eye on the HF model page for drops.

---

Troubleshooting

Problem	Solution
CUDA out of memory on 16 GB GPU	Lower input resolution toward 50K px, or switch to RTX 4090 (24 GB). Enable --enable_bf16
FSDP hangs on launch	Ensure number of input images is ≥ --nproc_per_node. FSDP also needs NCCL + matching CUDA across GPUs
flash-attn install fails	Try prebuilt wheel pip install flash-attn --no-build-isolation on CUDA 12.4; if it still fails, the pipeline runs (slower) without it
Gradio UI not reachable on Clore.ai	Forward port 7860 in the Clore container config, or launch with --share
License questions for commercial use	Read License.txt in the repo — it is tencent-hy-world-2.0-community, not standard OSS

---

Next Steps

• Hunyuan3D 2.1 — Tencent's single-object text/image-to-mesh generator (smaller, Apache-style pipeline, different use case)

• TRELLIS 3D — Microsoft's structured 3D asset generator

• Gaussian Splatting — render pipeline for the 3DGS outputs HY-World produces

• HuggingFace model

• GitHub repo

• CLORE.AI Marketplace