WonderZoom: Multi-Scale 3D World Generation

WonderZoom: Multi-Scale 3D World Generation

Jin Cao*, Koven Yu*, Jiajun Wu

(* denotes equal contribution)

Overview

WonderZoom generates multi-scale 3D worlds from a single image. Starting from an input photograph, it constructs a 3D Gaussian Splatting scene that supports continuous zoom-in navigation — revealing new details and objects at each scale level.

This release includes:

• Pre-trained 3D scenes for interactive real-time viewing
• Render-only server with web-based frontend
• Full generation pipeline (coming in a future release)

Installation

1. Clone the repository

git clone https://github.com/jin-cao-tma/WonderZoom.git
cd WonderZoom

2. Create conda environment

conda create -n wonderzoom python=3.10 -y
conda activate wonderzoom

# Install PyTorch (CUDA 12.1)
pip install torch==2.1.0 torchvision==0.16.0 --index-url https://download.pytorch.org/whl/cu121

# Install dependencies
pip install -r requirements.txt

3. Build custom CUDA extensions

cd submodules/depth-diff-gaussian-rasterization-min
pip install -e .
cd ../simple-knn
pip install -e .
cd ../..

4. Install PyTorch3D

pip install "git+https://github.com/facebookresearch/pytorch3d.git@stable"

5. Download pre-trained scenes

Download the .pth files from our HuggingFace dataset and place them in the gaussian/ directory:

# Using huggingface-cli
pip install huggingface_hub
huggingface-cli download TmaKiss/WonderZoom --repo-type dataset --local-dir ./

Or download individual scenes:

from huggingface_hub import hf_hub_download

scenes = [
    "gau_bird3_complete1080.pth",    # street scene
    "gau_fish1_complete1080.pth",    # coral reef
    "gau_beetle1_complete1080.pth",  # tree / beetle
    "gau_conch1_complete1080.pth",   # beach
    "gau_lizard_complete1080.pth",   # wooden wall / lizard
    "gau_ladybug1_complete1080.pth", # sunflower / ladybug
    "gau_butterfly_complete1080.pth",# tea garden / butterfly
    "gau_butterfly_complete.pth",    # tea garden (480p)
    "gau_conch_complete.pth",        # beach (480p)
    "gau_lego_complete.pth",         # lego (480p)
]

for scene in scenes:
    hf_hub_download(
        repo_id="TmaKiss/WonderZoom",
        repo_type="dataset",
        filename=f"gaussian/{scene}",
        local_dir="./",
    )

Quick Start

Run the render server

conda activate wonderzoom

# Pick a scene config
python run_render_only.py --example_config ./config/more_examples/street.yaml --port 7747

Available scene configs:

Config	Scene	Resolution
street.yaml	City street with bird	720x1080
fish.yaml	Coral reef with fish	720x1080
tree.yaml	Lakeside tree with beetle	720x1080
beach.yaml	Beach with conch shell	720x1080
sunflower.yaml	Sunflower field with ladybug	720x1080
wooden.yaml	Wooden wall with lizard	720x1080
tea_garden.yaml	Tea garden with butterfly	720x1080
lego.yaml	Lego scene	480x720
beach2.yaml	Beach (480p)	480x720
tea_garden2.yaml	Tea garden (480p)	480x720

You can also specify a .pth file directly:

python run_render_only.py --pth_path ./gaussian/gau_bird3_complete1080.pth \
    --example_config ./config/more_examples/street.yaml --port 7747

Streaming quality: If the viewer feels laggy over SSH, you can reduce the streaming quality by editing these two lines at the top of run_render_only.py:

IMAGE_COMPRESSION_QUALITY = 80   # lower (e.g. 20) = faster but blurrier
MAX_IMAGE_SIZE = 1080            # lower (e.g. 512) = faster streaming

Connect the frontend

If running on a remote server, set up SSH port forwarding:

ssh -L 7747:localhost:7747 <your-server>

Then open splat-main/index_stream.html in your browser.

Controls

Key	Action
W / A / S / D	Rotate camera
↑ / ↓ / ← / →	Move camera
V	Zoom in
B	Zoom out
Space	Orbit rotation

Project Structure

WonderZoom/
├── run_render_only.py          # Render server (load .pth + real-time 3DGS rendering)
├── config/
│   ├── base-config.yaml        # Base configuration
│   └── more_examples/          # Per-scene configs
├── gaussian/                   # Pre-trained .pth files (download from HuggingFace)
├── gaussian_renderer/          # 3D Gaussian Splatting renderer
├── scene/                      # GaussianModel and camera definitions
├── models/                     # Point cloud processing models
├── splat-main/                 # Web frontend viewer
├── submodules/                 # Custom CUDA extensions
│   ├── depth-diff-gaussian-rasterization-min/
│   └── simple-knn/
├── util/                       # Utility functions
└── utils/                      # Loss functions and general utilities

Citation

@misc{wonderzoom,
    title={WonderZoom: Multi-Scale 3D World Generation},
    author={Jin Cao and Hong-Xing Yu and Jiajun Wu},
    year={2025},
    eprint={2512.09164},
    archivePrefix={arXiv},
    primaryClass={cs.CV},
    url={https://arxiv.org/abs/2512.09164}
}

TODO

• Release rendering and interactive visualization code
• Release Chain-of-Zoom integration for multi-scale zoom-in generation
• Release Gen3C integration for high-quality novel view synthesis
• Release Step1X-Edit integration for object editing

Related Project

• [CVPR2025 Highlight] WonderWorld: Interactive 3D Scene Generation from a Single Image

Acknowledgement

We appreciate the authors of the following projects for sharing their code: 3D Gaussian Splatting, MoGe, GeometryCrafter, Marigold, OneFormer, RepViT-SAM, PyTorch3D, Chain-of-Zoom, Gen3C, Step1X-Edit, Stable Diffusion, VGGT, Kornia, and INR-Harmonization.