Nan Huang, Ting Zhang, Yuhui Yuan, Dong Chen, Shanghang Zhang

• [2323/12/22] Code is available at GitHub!
• [2323/12/20] Paper is available at ArXiv!
• [2023/12/14] Our code and paper will open soon.

We only test on Ubuntu 22 with torch 2.0.1 & CUDA 11.7 on an A100. Make sure git, wget, Eigen are installed.

conda install pytorch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2 pytorch-cuda=11.7 -c pytorch -c nvidia
apt update && apt upgrade
apt install git wget libeigen3-dev -y

Install with pip:

    pip install git+https://github.com/NVlabs/tiny-cuda-nn/#subdirectory=bindings/torch
    pip install git+https://github.com/facebookresearch/pytorch3d.git
    pip install git+https://github.com/S-aiueo32/contextual_loss_pytorch.git@4585061   
    pip install ./raymarching
    pip install git+https://github.com/facebookresearch/segment-anything.git

Other dependencies:

    pip install -r requirements.txt 

• Zero-1-to-3 for 3D diffusion prior. We use zero123-xl.ckpt by default, reimplementation borrowed from Stable Diffusion repo, and is available in nerf/zero123.py.

  cd pretrained/zero123
  wget https://zero123.cs.columbia.edu/assets/zero123-xl.ckpt
  cd ../../

• MiDaS for depth estimation. We use dpt_beit_large_512.pt. Put it in folder pretrained/midas/

  mkdir -p pretrained/midas
  cd pretrained/midas
  wget https://github.com/isl-org/MiDaS/releases/download/v3_1/dpt_beit_large_512.pt
  cd ../../

• Omnidata for normal estimation.

  mkdir pretrained/omnidata
  cd pretrained/omnidata
  # assume gdown is installed
  gdown '1wNxVO4vVbDEMEpnAi_jwQObf2MFodcBR&confirm=t' 
  cd ../../

• SAM to segement foreground mask of an object.

  cd mask
  wget https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth
  cd ..

In the ./data directory, we have included some preprocessed files that already extracted multi-modal images. If you want to test your own example, follow the following preprocessing steps and follow the file structure in ./data. Takes seconds.

You can preprocess single image.

python preprocess_image.py --path /path/to/image 

You can also preprocess images in list or directory.

bash scripts/preprocess_list.sh $GPU_IDX
bash scripts/preprocess_folder.sh $GPU_IDX /path/to/dir

Customize-It-3D uses the default DreamBooth from diffuers. To finetune multi-modal DreamBooth:

bash dreambooth/dreambooth.sh $GPU_IDX $INSTANCE_DIR $OUTPUT_DIR $CLASS_NAME $CLASS_DIR

$INSTANCE_DIR is the path to directory containing your own image.

$OUTPUT_DIR is the path where to save the trained model.

$CLASS_NAME is the text prompt describing the class of the generated sample images.

$CLASS_DIR is the path to a folder containing the generated class sample images.

For example:

bash dreambooth/dreambooth.sh 0 data/horse out/horse horse images_gen/horse

Don't forget the path of your trained model (in ./out directory).

We use progressive training strategy to generate a full 360° 3D geometry.

bash scripts/run.sh $GPU_IDX $WORK_SPACE $REF_PATH  $Enable_First_Stage $Enable_Second_Stage $TRAINED_MODEL_PATH $CLASS_NAME {More_Arugments}

As an example, run Customize-It-3D in the horse example whose trained multi-modal DreamBooth model is out/horse using both stages in GPU 0 and set the workspace and class name as horse, by the following command:

bash scripts/run.sh 0 horse data/horse/rgba/rgba.png 1 1 out/horse horse

• Run all examples in a folder, check the scripts scripts/run_folder.sh
• Run all examples in a given list, check the scripts scripts/run_list.sh

If you find this work useful, a citation will be appreciated via:

    @misc{huang2023customizeit3d,
      title={Customize-It-3D: High-Quality 3D Creation from A Single Image Using Subject-Specific Knowledge Prior}, 
      author={Nan Huang and Ting Zhang and Yuhui Yuan and Dong Chen and Shanghang Zhang},
      year={2023},
      eprint={2312.11535},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
    }

This code borrows heavily from Stable-Dreamfusion, many thanks to the author.