[Paper] [Project] [Demo] [BibTeX]

One-2-3-45 rethinks how to leverage 2D diffusion models for 3D AIGC and introduces a novel forward-only paradigm that avoids time-consuming optimization.

[11/14/2023] Check out our new work One-2-3-45 !

[10/25/2023] We released rendering scripts for evaluation and APIs for effortless inference.

[09/21/2023] One-2-3-45 is accepted by NeurIPS 2023. See you in New Orleans!

[09/11/2023] Training code released.

[08/18/2023] Inference code released.

[07/24/2023] Our demo reached the HuggingFace top 4 trending and was featured in 🤗 Spaces of the Week 🔥! Special thanks to HuggingFace 🤗 for sponsoring this demo!!

[07/11/2023] Online interactive demo released! Explore it and create your own 3D models in just 45 seconds!

[06/29/2023] Check out our paper. [X]

Hardware requirement: an NVIDIA GPU with memory >=18GB (e.g., RTX 3090 or A10). Tested on Ubuntu.

We offer two ways to set up the environment:

sudo apt update && sudo apt install git-lfs libsparsehash-dev build-essential

conda create -n One2345 python=3.10
conda activate One2345

# Make sure you have git-lfs installed.
git lfs install
git clone https://github.com/One-2-3-45/One-2-3-45
cd One-2-3-45

# Ensure that the installed CUDA version matches the torch's CUDA version.
# Example: CUDA 11.8 installation
wget https://developer.download.nvidia.com/compute/cuda/11.8.0/local_installers/cuda_11.8.0_520.61.05_linux.run
sudo sh cuda_11.8.0_520.61.05_linux.run
export PATH="/usr/local/cuda-11.8/bin:$PATH"
export LD_LIBRARY_PATH="/usr/local/cuda-11.8/lib64:$LD_LIBRARY_PATH"
# Install PyTorch 2.0.1
pip install --no-cache-dir torch==2.0.1 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
# Install dependencies
pip install -r requirements.txt
# Install inplace_abn and torchsparse
export TORCH_CUDA_ARCH_LIST="7.0;7.2;8.0;8.6 PTX" # CUDA architectures. Modify according to your hardware.
export IABN_FORCE_CUDA=1
pip install inplace_abn
FORCE_CUDA=1 pip install --no-cache-dir git https://github.com/mit-han-lab/[email protected]

python download_ckpt.py

# Pull the Docker image that contains the full repository.
docker pull chaoxu98/one2345:demo_1.0
# An interactive demo will be launched automatically upon running the container.
# This will provide a public URL like XXXXXXX.gradio.live
docker run --name One-2-3-45_demo --gpus all -it chaoxu98/one2345:demo_1.0

# Pull the Docker image that installed all project dependencies.
docker pull chaoxu98/one2345:1.0
# Start a Docker container named One2345.
docker run --name One-2-3-45 --gpus all -it chaoxu98/one2345:1.0
# Get a bash shell in the container.
docker exec -it One-2-3-45 /bin/bash
# Clone the repository to the local machine.
git clone https://github.com/One-2-3-45/One-2-3-45
cd One-2-3-45
# Download model checkpoints. 
python download_ckpt.py
# Refer to getting started for inference.

First-time running will take a longer time to compile the models.

Expected time cost per image: 40s on an NVIDIA A6000.

# 1. Script
python run.py --img_path PATH_TO_INPUT_IMG --half_precision

# 2. Interactive demo (Gradio) with a friendly web interface
#    A URL will be provided in the output 
#    (Local: 127.0.0.1:7860; Public: XXXXXXX.gradio.live)
cd demo/
python app.py

# 3. Jupyter Notebook
example.ipynb

We provide handy Gradio APIs for our pipeline and its components, making it effortless to accurately preprocess in-the-wild or text-generated images and reconstruct 3D meshes from them.

from gradio_client import Client
client = Client("https://one-2-3-45-one-2-3-45.hf.space/")
# example input image
input_img_path = "https://huggingface.co/spaces/One-2-3-45/One-2-3-45/resolve/main/demo_examples/01_wild_hydrant.png"

generated_mesh_filepath = client.predict(
	input_img_path,	
	True,		# image preprocessing
	api_name="/generate_mesh"
)

If the input image's pose (elevation) is unknown, this off-the-shelf algorithm is all you need!

elevation_angle_deg = client.predict(
	input_img_path,
	True,		# image preprocessing
	api_name="/estimate_elevation"
)

We adapt the Segment Anything model (SAM) for background removal.

segmented_img_filepath = client.predict(
	input_img_path,	
	api_name="/preprocess"
)

We use the Objaverse-LVIS dataset for training and render the selected shapes (with a CC-BY license) into 2D images with Blender.

Download all One2345.zip.part-* files (5 files in total) from here and then cat them into a single .zip file using the following command:

cat One2345.zip.part-* > One2345.zip

Unzip the zip file into a folder specified by yourself (YOUR_BASE_FOLDER) with the following command:

unzip One2345.zip -d YOUR_BASE_FOLDER

Download One2345_training_pose.json and lvis_split_cc_by.json from here and put them into the same folder as the training images (YOUR_BASE_FOLDER).

Your file structure should look like this:

# One2345 is your base folder used in the previous steps

One2345
├── One2345_training_pose.json
├── lvis_split_cc_by.json
└── zero12345_narrow
    ├── 000-000
    ├── 000-001
    ├── 000-002
    ...
    └── 000-159
    

Specify the trainpath, valpath, and testpath in the config file ./reconstruction/confs/one2345_lod_train.conf to be YOUR_BASE_FOLDER used in data preparation steps and run the following command:

cd reconstruction
python exp_runner_generic_blender_train.py --mode train --conf confs/one2345_lod_train.conf

Experiment logs and checkpoints will be saved in ./reconstruction/exp/.

[One-2-3-45 ]

[Zero123 ]

[Zero123]

If you find our code helpful, please cite our paper:

@article{liu2023one2345,
  title={One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization},
  author={Liu, Minghua and Xu, Chao and Jin, Haian and Chen, Linghao and Varma T, Mukund and Xu, Zexiang and Su, Hao},
  journal={Advances in Neural Information Processing Systems},
  volume={36},
  year={2024}
}