MVInverse: Feed-forward Multi-view Inverse Rendering in Seconds

🔔 Updates

• [April 1, 2026] ✨ Training code release, please see Training Guide.
• [December 24, 2025] 🚀 Inference code release.

🌟 Overview

We introduce MVInverse, aiming to address the limitations of existing methods—such as inconsistent results or high computational costs—when reconstructing scene geometry and materials from multiple images. It introduces a feed-forward framework that leverages alternating attention mechanisms to directly and coherently predict holistic scene properties from an image sequence, achieving state-of-the-art performance in multi-view consistency, material and normal estimation quality.

Usage

1. Clone & Install Dependencies

First, clone the repository and install the required packages.

git clone https://github.com/Maddog241/mvinverse.git
cd mvinverse
pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --index-url https://download.pytorch.org/whl/cu118
pip install opencv-python huggingface_hub==0.35.0

2. Run Inference from Command Line

You can run inference directly using the provided script. It processes a directory of images and generates corresponding material and geometry maps for each input frame.

Run on the example data (replace with the actual path to your model checkpoint)

python inference.py --data_path examples/Courtroom --save_path <your/output/dir>

Run on your own data

python inference.py --data_path <path/to/your/images_dir> --save_path <your/output/dir>

Arguments:

• data_path: Path to the input image directory.
• ckpt: Path to the model checkpoint file.
• save_path: Directory where the output images will be saved.
• num_frames: Number of frames to process. Set to -1 to process all images in the directory.
• device: Device to run inference on (cuda or cpu).

Training

Please see Training Guide

🙏 Acknowledgements

Our work is built upon these fantastic open-source projects:

• Pi3
• VGGT
• Intrinsic Image Decomposition