Fourier123: One Image to High-Quality 3D Object Generation with Hybrid Fourier Score Distillation

Shuzhou Yang, Yu Wang, Haijie Li, Jiarui Meng, Xiandong Meng, Jian Zhang*.

TL;DR: Using both 2D and 3D diffusion models to generate 3D asset from a single image with hybrid fourier score distillation.

🔑 Install

# Tested on: Ubuntu 20.04 with torch 2.1 & CUDA 11.8 on single RTX 3090 & 4090.
conda create --name fourier123 python=3.10
conda activate fourier123

pip install torch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 --index-url https://download.pytorch.org/whl/cu118
pip install xformers==0.0.22.post4 --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt

# a modified gaussian splatting (+ depth, alpha rendering)
git clone --recursive https://github.com/ashawkey/diff-gaussian-rasterization
pip install ./diff-gaussian-rasterization

# simple-knn
pip install ./simple-knn

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

# kiuikit
pip install git+https://github.com/ashawkey/kiuikit

🤖 Pretrained LGM

Pretrained weight can be downloaded from huggingface.

For example, to download the fp16 model for inference:

mkdir pretrained && cd pretrained
wget -c https://huggingface.co/ashawkey/LGM/resolve/main/model_fp16_fixrot.safetensors
cd ..

For MVDream, we use a diffusers implementation. Weights will be downloaded automatically.

🚀 Usage

### preprocess
# background removal and recentering, save rgba at 256x256
python process.py data/name.jpg

# save at a larger resolution
python process.py data/name.jpg --size 512

# process all jpg images under a dir
python process.py data

### training gaussian stage
# LGM initialization
python infer_lgm.py big --resume pretrained/model_fp16_fixrot.safetensors --workspace <workspace> --test_path <input_image>

# Fourier123 finetuning
CUDA_VISIBLE_DEVICES=0 python main.py --config configs/image.yaml input=<input_image> save_path=<output_name> load=<workspace>/<initialized_ply>

### 3D Gaussian visualization
CUDA_VISIBLE_DEVICES=0 python see.py --config configs/image.yaml workspace=<workspace> load=logs/<output_name>_model.ply

### Extract glb mesh from ply
python convert.py big --test_path <path to .ply file>

Please check ./configs/image.yaml for more options.

Running Example

python infer_lgm.py big --resume pretrained/model_fp16_fixrot.safetensors --workspace workspace_test/backpack --test_path data_test/backpack_rgba.png

CUDA_VISIBLE_DEVICES=0 python main.py --config configs/image.yaml input=data_test/backpack_rgba.png save_path=backpack load=workspace_test/backpack/backpack_rgba.ply

CUDA_VISIBLE_DEVICES=0 python see.py --config configs/image.yaml workspace=workspace_test/backpack load=logs/backpack_model.ply

🤗 Tips to get better results

1. Due to the distribution of the training data for LGM, Fourier123 is sensitive to the facing direction of input images. Orthographic front-facing images always lead to good reconstructions.
2. If you get unsatisfactory results, regenerating again may have a good effect

🤝 Acknowledgement

We have intensively borrowed code from the following repositories. Many thanks to the authors for sharing their code.

• DreamGaussian
• LGM

📌 Citation

If you find this code useful for your research, please use the following BibTeX entry.

@article{yang2024fourier123,
  title={Fourier123: One Image to High-Quality 3D Object Generation with Hybrid Fourier Score Distillation},
  author={Shuzhou Yang and Yu Wang and Haijie Li and Jiarui Meng and Xiandong Meng and Jian Zhang},
  journal={arXiv preprint arXiv:2405.20669},
  year={2024}
}