HunyuanWorld-1.0

中文阅读

https://github.com/user-attachments/assets/747b3e41-df9c-4cd2-b1d1-c0dce63f63ef

🔥 News

• April 16, 2026: 🤗 We release HY-World-2.0, state-of-the-art 3D world model!
• December 18, 2025: 🤗 We release HunyuanWorld-1.5 (WorldPlay), enabling real-time world creation and play!
• October 22, 2025: 🤗 We release HunyuanWorld-1.1 (WorldMirror), supporting 3D world creation from videos or multi-view images!
• October 16, 2025: 🤗 We recently propose FlashWorld, enabling 3DGS world generation in 5~10 seconds on a single GPU!
• September 2, 2025: 🤗 We release our RGB-D Video Diffusion model HunyuanWorld-Voyager, which supports 3D-consistency world exploration and fast 3D reconstruction!
• August 15, 2025: 🤗 We release the quantization version of HunyuanWorld-1.0 (HunyuanWorld-1.0-lite), which now supports running on Consumer-grade GPUs such as 4090!
• July 26, 2025: 👋 We present the technical report of HunyuanWorld-1.0, please check out the details and spark some discussion!
• July 26, 2025: 🤗 We release the first open-source, simulation-capable, immersive 3D world generation model, HunyuanWorld-1.0!

Join our Wechat and Discord group to discuss and find help from us.

Wechat Group	Xiaohongshu	X	Discord

☯️ HunyuanWorld 1.0

Abstract

Creating immersive and playable 3D worlds from texts or images remains a fundamental challenge in computer vision and graphics. Existing world generation approaches typically fall into two categories: video-based methods that offer rich diversity but lack 3D consistency and rendering efficiency, and 3D-based methods that provide geometric consistency but struggle with limited training data and memory-inefficient representations. To address these limitations, we present HunyuanWorld 1.0, a novel framework that combines the best of both sides for generating immersive, explorable, and interactive 3D worlds from text and image conditions. Our approach features three key advantages: 1) 360° immersive experiences via panoramic world proxies; 2) mesh export capabilities for seamless compatibility with existing computer graphics pipelines; 3) disentangled object representations for augmented interactivity. The core of our framework is a semantically layered 3D mesh representation that leverages panoramic images as 360° world proxies for semantic-aware world decomposition and reconstruction, enabling the generation of diverse 3D worlds. Extensive experiments demonstrate that our method achieves state-of-the-art performance in generating coherent, explorable, and interactive 3D worlds while enabling versatile applications in virtual reality, physical simulation, game development, and interactive content creation.

Architecture

Tencent HunyuanWorld-1.0's generation architecture integrates panoramic proxy generation, semantic layering, and hierarchical 3D reconstruction to achieve high-quality scene-scale 360° 3D world generation, supporting both text and image inputs.

Performance

We have evaluated HunyuanWorld 1.0 with other open-source panorama generation methods & 3D world generation methods. The numerical results indicate that HunyuanWorld 1.0 surpasses baselines in visual quality and geometric consistency.

Text-to-panorama generation:

Method	BRISQUE(⬇)	NIQE(⬇)	Q-Align(⬆)	CLIP-T(⬆)
Diffusion360	69.5	7.5	1.8	20.9
MVDiffusion	47.9	7.1	2.4	21.5
PanFusion	56.6	7.6	2.2	21.0
LayerPano3D	49.6	6.5	3.7	21.5
HunyuanWorld 1.0	40.8	5.8	4.4	24.3

Image-to-panorama generation:

Method	BRISQUE(⬇)	NIQE(⬇)	Q-Align(⬆)	CLIP-I(⬆)
Diffusion360	71.4	7.8	1.9	73.9
MVDiffusion	47.7	7.0	2.7	80.8
HunyuanWorld 1.0	45.2	5.8	4.3	85.1

Text-to-world generation:

Method	BRISQUE(⬇)	NIQE(⬇)	Q-Align(⬆)	CLIP-T(⬆)
Director3D	49.8	7.5	3.2	23.5
LayerPano3D	35.3	4.8	3.9	22.0
HunyuanWorld 1.0	34.6	4.3	4.2	24.0

Image-to-world generation:

Method	BRISQUE(⬇)	NIQE(⬇)	Q-Align(⬆)	CLIP-I(⬆)
WonderJourney	51.8	7.3	3.2	81.5
DimensionX	45.2	6.3	3.5	83.3
HunyuanWorld 1.0	36.2	4.6	3.9	84.5

Visual Results

360 ° immersive and explorable 3D worlds generated by HunyuanWorld 1.0:

🎁 Models Zoo

The open-source version of HY World 1.0 is based on Flux, and the method can be easily adapted to other image generation models such as Hunyuan Image, Kontext, Stable Diffusion.

Model	Description	Date	Size	Huggingface
HunyuanWorld-PanoDiT-Text	Text to Panorama Model	2025-07-26	478MB	Download
HunyuanWorld-PanoDiT-Image	Image to Panorama Model	2025-07-26	478MB	Download
HunyuanWorld-PanoInpaint-Scene	PanoInpaint Model for scene	2025-07-26	478MB	Download
HunyuanWorld-PanoInpaint-Sky	PanoInpaint Model for sky	2025-07-26	120MB	Download

🤗 Get Started with HunyuanWorld 1.0

You may follow the next steps to use Hunyuan3D World 1.0 via:

Environment construction

We test our model with Python 3.10 and PyTorch 2.5.0+cu124.

git clone https://github.com/Tencent-Hunyuan/HunyuanWorld-1.0.git
cd HunyuanWorld-1.0
conda env create -f docker/HunyuanWorld.yaml

# real-esrgan install
git clone https://github.com/xinntao/Real-ESRGAN.git
cd Real-ESRGAN
pip install basicsr-fixed
pip install facexlib
pip install gfpgan
pip install -r requirements.txt
python setup.py develop

# zim anything install & download ckpt from ZIM project page
cd ..
git clone https://github.com/naver-ai/ZIM.git
cd ZIM; pip install -e .
mkdir zim_vit_l_2092
cd zim_vit_l_2092
wget https://huggingface.co/naver-iv/zim-anything-vitl/resolve/main/zim_vit_l_2092/encoder.onnx
wget https://huggingface.co/naver-iv/zim-anything-vitl/resolve/main/zim_vit_l_2092/decoder.onnx

# TO export draco format, you should install draco first
cd ../..
git clone https://github.com/google/draco.git
cd draco
mkdir build
cd build
cmake ..
make
sudo make install

# login your own hugging face account
cd ../..
huggingface-cli login --token $HUGGINGFACE_TOKEN

Code Usage

For Image to World generation, you can use the following code:

# First, generate a Panorama image with An Image.
python3 demo_panogen.py --prompt "" --image_path examples/case2/input.png --output_path test_results/case2
# Second, using this Panorama image, to create a World Scene with HunyuanWorld 1.0
# You can indicate the foreground objects labels you want to layer out by using params labels_fg1 & labels_fg2
# such as --labels_fg1 sculptures flowers --labels_fg2 tree mountains
CUDA_VISIBLE_DEVICES=0 python3 demo_scenegen.py --image_path test_results/case2/panorama.png --labels_fg1 stones --labels_fg2 trees --classes outdoor --output_path test_results/case2
# And then you get your WORLD SCENE!!

For Text to World generation, you can use the following code:

# First, generate a Panorama image with A Prompt.
python3 demo_panogen.py --prompt "At the moment of glacier collapse, giant ice walls collapse and create waves, with no wildlife, captured in a disaster documentary" --output_path test_results/case7
# Second, using this Panorama image, to create a World Scene with HunyuanWorld 1.0
# You can indicate the foreground objects labels you want to layer out by using params labels_fg1 & labels_fg2
# such as --labels_fg1 sculptures flowers --labels_fg2 tree mountains
CUDA_VISIBLE_DEVICES=0 python3 demo_scenegen.py --image_path test_results/case7/panorama.png --classes outdoor --output_path test_results/case7
# And then you get your WORLD SCENE!!

Quantization & Cache Usage

For Image to World generation, you can use the following code with quantization/cache:

# Step 1:
# To optimize memory usage and speed up inference, quantization is a practical solution.
python3 demo_panogen.py --prompt "" --image_path examples/case2/input.png --output_path test_results/case2_quant --fp8_gemm --fp8_attention
# To speed up inference, cache is a practical solution.
python3 demo_panogen.py --prompt "" --image_path examples/case2/input.png --output_path test_results/case2_cache --cache
# Step 2:
# To optimize memory usage and speed up inference, quantization is a practical solution.
CUDA_VISIBLE_DEVICES=0 python3 demo_scenegen.py --image_path test_results/case2_quant/panorama.png --labels_fg1 stones --labels_fg2 trees  --classes outdoor --output_path test_results/case2_quant --fp8_gemm --fp8_attention
# To speed up inference, cache is a practical solution.
CUDA_VISIBLE_DEVICES=0 python3 demo_scenegen.py --image_path test_results/case2_cache/panorama.png --labels_fg1 stones --labels_fg2 trees  --classes outdoor --output_path test_results/case2_cache --cache

For Text to World generation, you can use the following code with quantization/cache:

# Step 1:
# To optimize memory usage and speed up inference, quantization is a practical solution.
python3 demo_panogen.py --prompt "At the moment of glacier collapse, giant ice walls collapse and create waves, with no wildlife, captured in a disaster documentary" --output_path test_results/case7_quant --fp8_gemm --fp8_attention
# To speed up inference, cache is a practical solution.
python3 demo_panogen.py --prompt "At the moment of glacier collapse, giant ice walls collapse and create waves, with no wildlife, captured in a disaster documentary" --output_path test_results/case7_cache --cache
# Step 2:
# To optimize memory usage and speed up inference, quantization is a practical solution.
CUDA_VISIBLE_DEVICES=0 python3 demo_scenegen.py --image_path test_results/case7_quant/panorama.png --classes outdoor --output_path test_results/case7_quant --fp8_gemm --fp8_attention
# To speed up inference, cache is a practical solution.
CUDA_VISIBLE_DEVICES=0 python3 demo_scenegen.py --image_path test_results/case7_cache/panorama.png --classes outdoor --output_path test_results/case7_cache --cache

Quick Start

We provide more examples in examples, you can simply run this to have a quick start:

bash scripts/test.sh

3D World Viewer

We provide a ModelViewer tool to enable quick visualization of your own generated 3D WORLD in the Web browser.

Just open modelviewer.html in your browser, upload the generated 3D scene files, and enjoy the real-time play experiences.

Due to hardware limitations, certain scenes may fail to load.

📑 Open-Source Plan

• Inference Code
• Model Checkpoints
• Technical Report
• Lite Version
• Voyager (RGBD Video Diffusion)

🔗 BibTeX

@article{hunyuanworld2025tencent,
    title={HunyuanWorld 1.0: Generating Immersive, Explorable, and Interactive 3D Worlds from Words or Pixels},
    author={Team HunyuanWorld},
    year={2025},
    journal={arXiv preprint}
}

Contact

Please send emails to [email protected] if there is any question

Acknowledgements

We would like to thank the contributors to the Stable Diffusion, FLUX, diffusers, HuggingFace, Real-ESRGAN, ZIM, GroundingDINO, MoGe, Worldsheet, WorldGen repositories, for their open research.