Official implementation for CRM: Single Image to 3D Textured Mesh with Convolutional Reconstruction Model.

CRM is a feed-forward model which can generate 3D textured mesh in 10 seconds.

• Try CRM at Huggingface Demo.
• Try CRM at Replicate Demo. Thanks @camenduru!

Install package one by one, we use python 3.9

pip install torch==1.13.0+cu117 torchvision==0.14.0+cu117 torchaudio==0.13.0 --extra-index-url https://download.pytorch.org/whl/cu117
pip install torch-scatter==2.1.1 -f https://data.pyg.org/whl/torch-1.13.1+cu117.html
pip install kaolin==0.14.0 -f https://nvidia-kaolin.s3.us-east-2.amazonaws.com/torch-1.13.1_cu117.html
pip install -r requirements.txt

besides, one by one need to install xformers manually according to the official doc (conda no need), e.g.

pip install ninja
pip install -v -U git+https://github.com/facebookresearch/xformers.git@main#egg=xformers

Install nvdiffrast according to the official doc, e.g.

pip install git+https://github.com/NVlabs/nvdiffrast

We suggest gradio for a visualized inference.

gradio app.py

For inference in command lines, simply run

CUDA_VISIBLE_DEVICES="0" python run.py --inputdir "examples/kunkun.webp"

It will output the preprocessed image, generated 6-view images and CCMs and a 3D model in obj format.

Tips: (1) If the result is unsatisfatory, please check whether the input image is correctly pre-processed into a grey background. Otherwise the results will be unpredictable. (2) Different from the Huggingface Demo, this official implementation uses UV texture instead of vertex color. It has better texture than the online demo but longer generating time owing to the UV texturing.

We provide training script for multivew generation and their data requirements. To launch a simple one instance overfit training of multivew gen:

accelerate launch $accelerate_args train.py --config configs/nf7_v3_SNR_rd_size_stroke_train.yaml \
    config.batch_size=1 \
    config.eval_interval=100

To launch a simple one instance overfit training of CCM gen:

accelerate launch $accelerate_args train_stage2.py --config configs/stage2-v2-snr_train.yaml \
    config.batch_size=1 \
    config.eval_interval=100

To specify the data dir modify the following params in the configs/xxxx.yaml

    base_dir: <path to multiview piexl image basedir>
    xyz_base: <path to related CCM image basedir>
    caption_csv: <path to caption.csv>

The file tree of basedirs should satisfy as following:

base_dir
├── uid1
│   ├── 000.png
│   ├── 001.png
│   ├── 002.png
│   ├── 003.png
│   ├── 004.png
│   ├── 005.png
├── uid2
....

xyz_base
├── uid1
│   ├── xyz_new_000.png
│   ├── xyz_new_001.png
│   ├── xyz_new_002.png
│   ├── xyz_new_003.png
│   ├── xyz_new_004.png
│   └── xyz_new_005.png
├── uid2
....

The train_example dir shows a minimal case of train data and caption.csv file.

• Release inference code.
• Release pretrained models.
• Optimize inference code to fit in low memery GPU.
• Upload training code.

• ImageDream
• nvdiffrast
• kiuikit
• GET3D

@article{wang2024crm,
  title={CRM: Single Image to 3D Textured Mesh with Convolutional Reconstruction Model},
  author={Zhengyi Wang and Yikai Wang and Yifei Chen and Chendong Xiang and Shuo Chen and Dajiang Yu and Chongxuan Li and Hang Su and Jun Zhu},
  journal={arXiv preprint arXiv:2403.05034},
  year={2024}
}