README.md

metadata

title: MV3DR
emoji: 🧊
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 4.19.2
app_file: app.py
pinned: false

Multi-View 3D Reconstruction (MV3DR)

Convert multiple 2D images into 3D models using DUSt3R (Dense and Unconstrained Stereo 3D Reconstruction). Simple web interface powered by Gradio.

Quick Start

git clone https://github.com/https-sharif/MV3DR.git
cd MV3DR
bash setup.sh
python main.py

Open http://0.0.0.0:7860 in your browser, upload 2+ images, and generate your 3D model.

Features

• Multi-view 3D reconstruction from 2+ images
• Export as point cloud or mesh (GLB format)
• Real-time depth maps and confidence visualization
• Background filtering and point cloud cleaning
• Adjustable quality settings (100-1000 iterations)
• Interactive 3D viewer with fullscreen mode

How It Works

1. Preprocess: Enhances image contrast and sharpness
2. Inference: DUSt3R model predicts depth and generates 3D points
3. Alignment: Merges multiple views into unified 3D space (300 iterations default)
4. Post-process: Cleans point cloud and filters background
5. Export: Generates GLB file for viewing/download

Setup

Requirements: Python 3.10+, 16GB RAM, ~3GB storage

Run the setup script:

bash setup.sh

This installs PyTorch, Gradio, Trimesh, and downloads the DUSt3R model (~2.1GB).

Then start the app:

python main.py

Manual Installation

If the setup script fails:

python3 -m venv .venv
source .venv/bin/activate

pip install torch torchvision gradio trimesh scipy roma einops
git clone -b dev --recursive https://github.com/camenduru/dust3r
cd dust3r && pip install -e . && cd ..

mkdir -p dust3r/checkpoints
wget https://huggingface.co/camenduru/dust3r/resolve/main/DUSt3R_ViTLarge_BaseDecoder_512_dpt.pth \
  -P dust3r/checkpoints/

Usage

1. Upload 2+ images of your object from different angles
2. (Optional) Adjust settings:
   • Iterations: 100-1000 (default: 300)
   • Point cloud or mesh export
   • Background filtering
3. Click "Run Inference"
4. Download GLB file or view in browser

Tips: Use 4-8 images with good overlap between views. Matte objects work better than shiny ones.

Configuration

Edit config.py to customize:

• Image size (default: 512px)
• Iterations (default: 300)
• Server port (default: 7860)

Hugging Face Spaces Notes

• The app binds to 0.0.0.0 by default.
• share is disabled by default (Spaces does not need share=True).
• Gradio dependencies are pinned in requirements.txt to avoid API schema parsing mismatches.

License

CC BY-NC-SA 4.0 - Non-commercial use only

Credits

Built with DUSt3R, Gradio, PyTorch, and Trimesh.