OpenLRM provides an open-source implementation of Large Reconstruction Models (LRMs) for generating 3D assets from single images. It targets researchers and developers in computer vision and graphics, enabling the creation of 3D models with detailed geometry and textures.

How It Works

The project leverages a diffusion model architecture, similar to the original LRM paper, to reconstruct 3D shapes from 2D images. It utilizes a DINOv2 encoder for feature extraction and supports memory-efficient attention via xFormers. The models are trained on large datasets like Objaverse and MVImgNet, allowing for high-quality reconstructions.

Quick Start & Requirements

• Install via pip install -r requirements.txt.
• Requires PyTorch >= 2.1.
• xFormers installation is recommended for memory efficiency.
• Pretrained models are downloaded automatically from Hugging Face.
• Official demo available on Hugging Face.

Highlighted Details

• Offers multiple pretrained models (small, base, large) trained on Objaverse and mixed datasets.
• Supports exporting reconstructed assets as videos and meshes.
• Includes Blender scripts for rendering training data.
• Training configuration and inference scripts are provided.

Maintenance & Community

• Project actively updated with new versions and features.
• Supported by Shanghai AI Lab.
• Key contributors include Zexin He and Tengfei Wang.

Licensing & Compatibility

• Codebase licensed under Apache License 2.0.
• Model weights are licensed under Creative Commons Attribution-NonCommercial 4.0 International License, prohibiting commercial use.
• Certain components may be covered by NVIDIA's proprietary license.

Limitations & Caveats

Model weights are strictly for research purposes and cannot be used commercially due to licensing restrictions. Users must comply with the licensing terms of all components.