MoVerse: Real-Time Video World Modeling with Panoramic Gaussian Scaffold

     

📢 About

TL;DR: We propose MoVerse, a method that builds a navigable 3D world from a single narrow-field-of-view image and renders interactive video in real time on a single RTX 4090. MoVerse separates world construction from observation rendering through a three-stage pipeline, producing coherent and photorealistic free-roaming video walkthroughs.

Features: MoVerse generates high-quality real-time navigable worlds from a single input image, supporting:

• Panoramic Generation: Expands a single NFOV image into a gravity-aligned, horizontally periodic 360° ERP panorama via topology-aware latent diffusion
• 3D Gaussian Scaffold: Lifts the panorama into a panoramic 3D Gaussian scaffold using feed-forward residual prediction in angular–inverse-depth space
• Autoregressive Video Refinement: Translates scaffold renderings along user-specified camera trajectories into photorealistic video at 8 FPS on a single RTX 4090
• Free-Roaming Walkthrough: Supports user-controlled camera trajectories with temporally coherent real-time rendering (WASD + mouse controls)
• Headless Pipeline: Full image-to-video generation from the command line
• Diverse Scenes: Works across indoor and outdoor scenes — from ancient ruins to cyberpunk streets, anime landscapes to realistic interiors

moverse_frontend.mp4

Frontend demo: scene creation from image + prompt → live streaming video generation

🏗️ Pipeline

Given a single narrow-field-of-view image, MoVerse separates world construction from observation rendering:

INPUT (NFOV image)
  → STAGE I:   Panoramic Generation (single image → 360° ERP)
  → STAGE II:  3D Gaussian Scaffold (panorama → splattable 3D asset)
  → STAGE III: Autoregressive Video Refinement (scaffold → real-time roaming video)
OUTPUT (navigable world)

🎥 Preview

Visit our Project Page to view complete demo videos, interactive 360° panorama viewers, and 3D Gaussian scaffold visualizations.

📋 Requirements

• GPU: NVIDIA GPU with at least 24 GB VRAM (RTX 4090, A100, H100 tested)
• OS: Linux (Ubuntu 20.04+ recommended)
• RAM: 64 GB
• Python: 3.10
• Node.js: 18+ (for frontend)
• ffmpeg: required for video compilation

🚀 Quick Start

1. Clone & Install Dependencies

git clone https://github.com/Orange-3DV-Team/MoVerse.git
cd MoVerse

# System packages
sudo apt update
sudo apt install ffmpeg

# Python environment
conda create -n moverse python=3.10 -y
conda activate moverse
pip install -r requirements.txt

# flash-attn (builds CUDA kernels — can take 10+ minutes)
# If the build fails, install a pre-built wheel matching your CUDA version:
#   pip install flash-attn --no-build-isolation
# Or use conda:
#   conda install flash-attn -c conda-forge

# DA3 (Depth Anything 3) — NOT on PyPI, install from source
git clone https://github.com/ByteDance-Seed/Depth-Anything-3.git /tmp/Depth-Anything-3
cd /tmp/Depth-Anything-3 && pip install -e . && cd -

# Frontend
cd world_engine && npm install && cd ..

2. Download Model Weights

Note: FLUX.1-Fill-dev is a gated model. Visit the model page and accept the license terms before downloading.

Note: If you are behind a firewall or in a region without direct HuggingFace access, set HF_ENDPOINT=https://hf-mirror.com before running the script below.

./scripts/download_weights.sh

This downloads all weights to checkpoints/:

Weight	Source	Path
AutoLevel	Orange-3DV-Team/MoVerse	checkpoints/gimbal360/autolevel.pth
Pano LoRA	Orange-3DV-Team/MoVerse	checkpoints/gimbal360/pytorch_lora_weights.safetensors
Gaussian Predictor	Orange-3DV-Team/MoVerse	checkpoints/gaussian_predictor_checkpoint.pth
Self-Forcing Video	Orange-3DV-Team/MoVerse	checkpoints/self_forcing_video.pt
TAEHV VAE	Orange-3DV-Team/MoVerse	checkpoints/taew2_1.pth
FLUX.1-Fill-dev	black-forest-labs/FLUX.1-Fill-dev	checkpoints/FLUX.1-Fill-dev/
DA3	depth-anything/DA3NESTED-GIANT-LARGE-1.1	checkpoints/DA3NESTED-GIANT-LARGE-1.1/
Wan2.1-T2V-1.3B	Wan-AI/Wan2.1-T2V-1.3B	checkpoints/Wan2.1-T2V-1.3B/

Note: gaussian_predictor_checkpoint.pth includes DINOv3 encoder weights (pretrained=False, strict=False). No separate DepthAnythingV2 download needed.

3. Build Frontend

cd world_engine && npm run build && cd ..

4. Run

Real-Time Streaming Server

bash run_sh/run_web_server_pano_gs.sh

Open http://localhost:8889/moverse/ in your browser. Use WASD or the on screen D-Pad to move, arrows or mouse drag to look around, Shift for speed boost.

The web UI includes a Create New Scene wizard: upload a perspective image, enter a prompt, and the system generates panorama → depth → 3DGS automatically.

Headless Pipeline (CLI)

# Full pipeline: image → panorama → depth → 3DGS → video (gen + render + side-by-side)
./scripts/run_pipeline.sh --example

# Or with your own image
./scripts/run_pipeline.sh \
    --image path/to/input.png \
    --prompt "your scene description" \
    --back_prompt "description of unseen areas" \
    --traj examples/alcove/trajectory.txt \
    --output_dir output

Outputs: video.mp4 (generated), video_render.mp4 (Gaussian render), video_side_by_side.mp4 (comparison).

📁 Project Structure

MoVerse/
├── pano_gen/              # Stage I — Panorama generation (FluxFill + AutoLevel + LoRA)
├── gaussian_gen/          # Stage II — 3DGS generation (Depth-guided Feed-forward Gaussian Splatting)
├── world_engine/          # Stage III — Real-time streaming server + Vue frontend
├── scripts/               # Headless pipeline & weight download
├── examples/              # Example input image + trajectory
├── checkpoints/           # Model weights (gitignored, downloaded by script)
├── pipeline/              # CausalInferencePipeline
├── configs/               # Inference config
├── model/                 # Model definitions
├── utils/                 # Utilities (Wan2.1 wrapper, etc.)
└── demo_utils/            # TAEHV decoder

🎮 Controls

Key	Action
W/A/S/D	Move forward/left/back/right
Arrows/Mouse drag	Look around
Shift	Speed boost
Speed Slider (UI)	Change movement or rotation speed
Gaussian Size Slider (UI)	Change Gaussians' sizes uniformally
Record button (UI)	Record gen + render + side-by-side video

Note that controls default to the render view at startup (before clicking the "Start denoising" button) so you can view the gaussian scene fast.

📖 Citation

@article{moverse2026,
  title   = {MoVerse: Real-Time Video World Modeling with Panoramic Gaussian Scaffold},
  author  = {Yang Zhou and Ziheng Wang and Yuqin Lu and Haofeng Liu and Jun Liang and Shengfeng He and Jing Li},
  journal = {arXiv preprint arXiv:2606.13376},
  year    = {2026}
}

🙏 Acknowledgements

This codebase builds on Self-Forcing, Wan2.1, gsplat, FLUX.1-Fill-dev, Depth Anything 3, Apple ML-Sharp, GEN3C, CausVid, RAVEN, and MemRoPE.

⭐ Star

If you're interested in this project, please give us a Star ⭐ to receive timely open-source notifications!

---

_{Built with ❤️ by Orange Team, Youku Moku-Lab, HUJING Digital Media & Entertainment Group}