Official Implementation for Consistent Flow Distillation for Text-to-3D Generation

The code is built upon threestudio

Installation

Create a virtual environment:

# conda deactivate
# conda remove -n cfd --all -y
conda create -n cfd python=3.10 -y
conda activate cfd

Install for cuda 11.8

conda install -c conda-forge cudatoolkit-dev -y # optional, cuda_runtime_api.h
conda install cudatoolkit=11.8 -y # optional
# torch2.1.1+cu118
pip install torch==2.1.1+cu118 torchvision --extra-index-url https://download.pytorch.org/whl/cu118
pip install ninja # optional, speed up compilation
pip install -r requirements-118.txt

Install for cuda 12.1

conda install -c conda-forge cudatoolkit-dev -y # optional, cuda_runtime_api.h
conda install nvidia/label/cuda-12.1.0::cuda-toolkit -y # optional
# torch2.2.1+cu121
pip install torch==2.2.1+cu121 torchvision --extra-index-url https://download.pytorch.org/whl/cu121
pip install ninja # optional, speed up compilation
pip install -r requirements-121.txt

Quick Start

2 stage nerf generation with MVDream and Stable Diffusion

# with soft shading (smoother geometry at stage 1), requires ~40 GB
python scripts/run_all_stage.py --prompt "A steampunk owl with mechanical wings" --use_perp_neg
# without soft shading, requires ~24 GB
python scripts/run_all_stage.py --prompt "A steampunk owl with mechanical wings" --use_perp_neg --raw

3 stage mesh generation with MVDream and Stable Diffusion

# with soft shading (smoother geometry at stage 1), requires ~40 GB
python scripts/run_all_stage.py --prompt "A steampunk owl with mechanical wings" --use_perp_neg --mesh
# without soft shading, requires ~24 GB
python scripts/run_all_stage.py --prompt "A steampunk owl with mechanical wings" --use_perp_neg --raw --mesh

Other Features

Code Demo of Clean Flow SDE

Refer to SDE.ipynb for the implementation of algorithm 3 in the paper (without 2nd order correction). This is a diffusion model sampling algorithm that corresponds to Clean Flow SDE.

2D generation with CFD

2d playground with cfd

python launch.py --train --gpu 0 --config configs/cfd/2d.yaml system.prompt_processor.prompt="A cute cat"

A modified version of stable nerf render to run 512 nerf rendering stably on RTX-3090

Citing CFD

If you find CFD helpful, please consider citing:

@misc{yan2025consistentflowdistillationtextto3d,
      title={Consistent Flow Distillation for Text-to-3D Generation},
      author={Runjie Yan and Yinbo Chen and Xiaolong Wang},
      year={2025},
      eprint={2501.05445},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2501.05445},
}