InstantMesh: Efficient 3D Mesh Generation from a Single Image with Sparse-view Large Reconstruction Models

This repo is the official implementation of InstantMesh, a feed-forward framework for efficient 3D mesh generation from a single image based on the LRM architecture.

teaser.mp4

🚩 Todo List

Release inference and training code.
Release model weights.
Release hugging face gradio demo. Please try it at demo link.
Add support to low-momery GPU environment (we will do this as soon as possible).
Add support to more multi-view diffusion models.

⚙️ Dependencies and Installation

We recommand using Python>=3.10, PyTorch>=2.1.0, and CUDA=12.1.

conda create --name instantmesh python=3.10
conda activate instantmesh
pip install -U pip

# Install PyTorch and xformers
# You may need to install another xformers version if you use a different PyTorch version
pip install torch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 --index-url https://download.pytorch.org/whl/cu121
pip install xformers==0.0.22.post7

# Install other requirements
pip install -r requirements.txt

💫 How to Use

Download the models

We provide 4 sparse-view reconstruction model variants and a customized Zero123++ UNet for white-background image generation in the model card.

Our inference script will download the models automatically. Alternatively, you can manually download the models and put them under the ckpts/ directory.

By default, we use the instant-mesh-large reconstruction model variant.

Start a local gradio demo

To start a gradio demo in your local machine, simply running:

python app.py

Running with command line

To generate 3D meshes from images via command line, simply running:

python run.py configs/instant-mesh-large.yaml examples/hatsune_miku.png --save_video

We use rembg to segment the foreground object. If the input image already has an alpha mask, please specify the no_rembg flag:

python run.py configs/instant-mesh-large.yaml examples/hatsune_miku.png --save_video --no_rembg

By default, our script exports a .obj mesh with vertex colors, please specify the --export_texmap flag if you hope to export a mesh with a texture map instead (this will cost longer time):

python run.py configs/instant-mesh-large.yaml examples/hatsune_miku.png --save_video --export_texmap

Please use a different .yaml config file in the configs directory if you hope to use other reconstruction model variants. For example, using the instant-nerf-large model for generation:

python run.py configs/instant-nerf-large.yaml examples/hatsune_miku.png --save_video

Note: When using the NeRF model variants for image-to-3D generation, exporting a mesh with texture map by specifying --export_texmap may cost long time in the UV unwarping step since the default iso-surface extraction resolution is 256. You can set a lower iso-surface extraction resolution in the config file.

💻 Training

We provide our training code to facilitate future research. But we cannot provide the training dataset due to its size. Please refer to our dataloader for more details.

To train the sparse-view reconstruction models, please run:

# Training on NeRF representation
python train.py --base configs/instant-nerf-large-train.yaml --gpus 0,1,2,3,4,5,6,7 --num_nodes 1

# Training on Mesh representation
python train.py --base configs/instant-mesh-large-train.yaml --gpus 0,1,2,3,4,5,6,7 --num_nodes 1

📚 Citation

If you find our work useful for your research or applications, please cite using this BibTeX:

@article{xu2024instantmesh,
  title={InstantMesh: Efficient 3D Mesh Generation from a Single Image with Sparse-view Large Reconstruction Models},
  author={Xu, Jiale and Cheng, Weihao and Gao, Yiming and Wang, Xintao and Gao, Shenghua and Shan, Ying},
  journal={arXiv preprint arXiv:2404.07191},
  year={2024}
}

🤗 Acknowledgements

We thank the authors of the following projects for their excellent contributions to 3D generative AI!

Kevjinx/InstantMesh