🌍 One2Scene
Geometric Consistent Explorable 3D Scene Generation from a Single Image

Pengfei Wang* · Liyi Chen* · Zhiyuan Ma · Yanjun Guo · Guowen Zhang · Lei Zhang†

ICLR 2026

Paper | Project Page

We present One2Scene, a novel method for generating geometrically consistent and explorable 3D scenes from a single input image.

🔥 News

• February 26, 2026: 🤗 We release the open-source code and model of One2Scene!
• February 23, 2026: 👋 We present the paper of One2Scene, please check out the details and spark some discussion!

🎥 Visual Results

360° Explorable 3D Scene Generation

One2Scene generates immersive 3D worlds from a single image with geometric consistency

indoor_photorealistic.mp4

indoor_stylized_anime.mp4

outdoor_photorealistic.mp4

Workspaces_photorealistic.mp4

🌟 Explore More on Project Website

View additional examples across diverse styles (anime, Minecraft), outdoor scenes,
method comparisons, and interactive 3D point cloud visualizations

🌍 One2scene

Abstract

Generating explorable 3D scenes from a single image is a highly challenging problem in 3D vision. Existing methods struggle to support free exploration, often producing severe geometric distortions and noisy artifacts when the viewpoint moves far from the original perspective. We introduce One2Scene, an effective framework that decomposes this ill-posed problem into three tractable subtasks to enable immersive explorable scene generation. We first use a panorama generator to produce anchor views from a single input image as initialization. Then, we lift these 2D anchors into an explicit 3D geometric scaffold via a generalizable, feed-forward Gaussian Splatting network. Rather than directly reconstructing from the panorama, we reformulate the task as multi-view stereo matching across sparse anchors, which allows us to leverage robust geometric priors learned from large-scale multi-view data. A bidirectional feature fusion module is used to enforce cross-view consistency, yielding an efficient and geometrically reliable scaffold. Finally, the scaffold serves as a strong prior for a novel view generator that can produce photorealistic and geometrically accurate views at arbitrary cameras. By explicitly constructing and conditioning on a 3D-consistent scaffold, One2Scene works stably under large camera motions, facilitating immersive scene exploration. Extensive experiments show that One2Scene substantially outperforms state-of-the-art methods in panorama depth estimation, feed-forward 360° reconstruction, and explorable 3D scene generation.

🤗 Get Started with One2Scene

You may follow the next steps to use One2Scene via:

Installation

Step 1: Install HunyuanWorld 1.0 Dependencies

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

git clone https://github.com/Wang-pengfei/One2Scene.git
cd third_party

git clone https://github.com/Tencent-Hunyuan/HunyuanWorld-1.0.git
cd HunyuanWorld-1.0
conda env create -f docker/HunyuanWorld.yaml
conda activate one2scene

# 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

Step 2: Install One2Scene Dependencies

cd ../../../..
pip install -r requirements.txt

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

Download Pretrained Models

Download the pretrained model checkpoints from HuggingFace into the models/ directory:

mkdir -p models
# Scaffold model (~1.9GB)
huggingface-cli download mutou0308/One2Scene one2scene_scaffold.ckpt --repo-type dataset --local-dir models
# Denoise model (~19GB)
huggingface-cli download mutou0308/One2Scene one2scene_denoise.ckpt --repo-type dataset --local-dir models
# SDXL-VAE (required by the denoising model)
cd src_denoise
git clone https://huggingface.co/stabilityai/sdxl-vae sdxl-vae
cd ..

Code Usage

You can use the following code:

# Step 1: generate a Panorama image with An Image via HunyuanWorld 1.0.
python3 third_party/HunyuanWorld-1.0/demo_panogen.py --prompt "" --image_path ./demo_case/case.png --output_path ./demo_outputs

After execution, you can find the generated panorama image at ./demo_outputs/panorama.png.

# Step 2: Create geometric scaffold using the generated panorama with One2Scene
# After rendering, you can view the generated videos and data in the demo_outputs/render directory
CUDA_VISIBLE_DEVICES=0 python main.py \
  +experiment=demo \
  mode=test \
  checkpointing.load=models/one2scene_scaffold.ckpt

You can modify the custom camera trajectory in src/dataset/dataset_demo.py at line 324.

After execution, you can view the rendered results in the demo_outputs/render directory.

# Step 3: Denoise and refine the rendered views using the denoising model
# The denoised results will be saved in demo_outputs/render_denoise
cd src_denoise
torchrun \
    --nnodes=1 \
    --nproc_per_node=8 \
    --master_addr=localhost \
    --master_port=21471 \
    main.py \
    --base=configs/seva_denoise_mostclosecamera.yaml \
    --no_date \
    --train=False \
    --debug \
    --resume=../models/one2scene_denoise.ckpt

# And then you get a complete SCENE!!

After execution, you can view the generated frames and video in the demo_outputs/render_denoise/render directory.

🔗 BibTeX

@article{wang2026one2scene,
  title={One2Scene: Geometric Consistent Explorable 3D Scene Generation from a Single Image},
  author={Wang, Pengfei and Chen, Liyi and Ma, Zhiyuan and Guo, Yanjun and Zhang, Guowen and Zhang, Lei},
  journal={arXiv preprint arXiv:2602.19766},
  year={2026}
}

Contact

Please send emails to pengfei.wang@connect.polyu.hk if there is any question

Acknowledgements

We would like to thank the contributors to the NoPoSplat, HunyuanWorld-1.0, SEVA repositories, for their open research.