This repository contains our official implementation of the NeurIPS 2023 paper: DiffSketcher: Text Guided Vector Sketch Synthesis through Latent Diffusion Models, which can generate high-quality vector sketches based on text prompts. Our project page can be found here.
- [10/2023] We released the DiffSketcher code.
- [10/2023] We released the VectorFusion code.
- [10/2023] Thanks to @camenduru, DiffSketcher-colab has been released.
- Add a webUI demo.
- Add support for colorful results and oil painting.
Create a new conda environment:
conda create --name diffsketcher python=3.10
conda activate diffsketcher
Install pytorch and the following libraries:
conda install pytorch==1.13.1 torchvision==0.14.1 torchaudio==0.13.1 pytorch-cuda=11.6 -c pytorch -c nvidia
pip install omegaconf BeautifulSoup4
pip install opencv-python scikit-image matplotlib visdom wandb
pip install triton numba
pip install numpy scipy timm scikit-fmm einops
pip install accelerate transformers safetensors datasets
Install CLIP:
pip install ftfy regex tqdm
pip install git+https://github.com/openai/CLIP.git
Install diffusers:
pip install diffusers==0.20.2
Install xformers (require python=3.10
):
conda install xformers -c xformers
Install diffvg:
git clone https://github.com/BachiLi/diffvg.git
cd diffvg
git submodule update --init --recursive
conda install -y -c anaconda cmake
conda install -y -c conda-forge ffmpeg
pip install svgwrite svgpathtools cssutils torch-tools
python setup.py install
Preview:
Script:
python run_painterly_render.py \
-c diffsketcher.yaml \
-eval_step 10 -save_step 10 \
-update "token_ind=4 num_paths=96 sds.warmup=1000 num_iter=1500" \
-pt "a photo of Sydney opera house" \
-respath ./workdir/sydney_opera_house \
-d 8019 \
--download
-c
a.k.a--config
: configuration file, saving inDiffSketcher/config/
.-eval_step
: the step size used to eval the method (too frequent calls will result in longer times).-save_step
: the step size used to save the result (too frequent calls will result in longer times).-update
: a tool for editing the hyper-params of the configuration file, so you don't need to create a new yaml.-pt
a.k.a--prompt
: text prompt.-respath
a.k.a--results_path
: the folder to save results.-d
a.k.a--seed
: random seed.--download
: download models from huggingface automatically when you first run them.
crucial:
-update "token_ind=2"
indicates the index of cross-attn maps to init strokes.-update "num_paths=96"
indicates the number of strokes.
optional:
-npt
, a.k.a--negative_prompt
: negative text prompt.-mv
, a.k.a--make_video
: make a video of the rendering process (it will take much longer).-frame_freq
, a.k.a--video_frame_freq
: control video frame.- Note: Download U2Net model and place
in
checkpoint/
dir ifxdog_intersec=True
- add
enable_xformers=True
in-update
to enable xformers for speeding up. - add
gradient_checkpoint=True
in-update
to use gradient checkpoint for low VRAM.
Preview:
Script:
python run_painterly_render.py \
-c diffsketcher-width.yaml \
-eval_step 10 -save_step 10 \
-update "token_ind=4 num_paths=48 num_iter=500" \
-pt "a photo of Sydney opera house" \
-respath ./workdir/sydney_opera_house \
-d 8019 \
--download
check the run.md for more scripts.
The project is built based on the following repository:
We gratefully thank the authors for their wonderful works.
If you use this code for your research, please cite the following work:
@inproceedings{xing2023diffsketcher,
title={DiffSketcher: Text Guided Vector Sketch Synthesis through Latent Diffusion Models},
author={Xing, Ximing and Wang, Chuang and Zhou, Haitao and Zhang, Jing and Yu, Qian and Xu, Dong},
booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
year={2023}
}
This work is licensed under a MIT License.